Anthropic is making Claude Code’s auto mode the default for its paid and enterprise users, allowing the coding agent to execute more actions without requiring developers to approve each one.

“Starting August 14, 2026, auto mode becomes the default permission mode for new sessions on Pro, Max, and Team plans,” the company wrote in the coding agent’s documentation, adding that the same change is planned for Claude Enterprise, API, and cloud platform users within the next month.

That essentially means developers using those plans will no longer have to manually approve every tool call or action Claude Code wants to make while executing a task.

Instead, each tool call is evaluated by an automated classifier designed to determine whether the action is safe to execute. Actions considered irreversible, destructive, or outside the agent’s environment can still be blocked, with Claude Code either attempting a safer approach or asking the developer for approval, the company wrote in a blog post.

“If it can’t make progress — three blocks in a row, or twenty across a session — Claude Code falls back to manual approvals,” it explained.

This reduction in manual intervention, Anthropic further added, is intended to make the coding agent more suitable for long-running tasks as well as cut down on “permission fatigue” that it says has been threatening to reduce its security posture.

According to its internal data, Claude Code users typically approve 97% of permission prompts, while only 3% are rejected. It also found that 49.5% of active CLI users had created a Bash allow-rule by June, while 62% had used “bypassPermissions” or selected “don’t ask again” for Bash.

In contrast, the default auto mode, in an Anthropic-controlled study involving 1,053 paid professional testers, caught 89% of deliberate dangerous commands compared to just 13.6% being detected by human reviewers.

More productivity but with tradeoffs

Analysts too see auto mode as a step toward making coding agents more autonomous, but with tradeoffs.

“Developers should see fewer permission pop-ups and spend less time watching Claude, meaning they can give it larger tasks and let it edit files, run tests, and execute routine commands with fewer interruptions. That way, developers can finish and ship more work using this mode because they aren’t constantly interrupted by approval windows,” said Pareekh Jain, principal analyst at Pareekh Consulting.

However, that same classifier in auto mode that reduces permission pop-ups could also result in more latency for routine or quick tasks, as trusted workflows also pass through it, adding an extra execution step, noted Manoj Chandra Jha, principal analyst at Nord-IQ Research.

Another tradeoff of reducing real-time approvals, according to Amit Kumar Jena, AI development manager at IT consulting firm Kanerika, is that it could make it harder for enterprises to maintain effective human oversight as agents take on longer-running tasks. “Reviewing four hours of unattended agent output is a harder skill, and most teams haven’t built it,” Jena said.

Better governance for CIOs

For CIOs, though, auto mode can strengthen governance in software development.

Rather than policing individual developer actions, CIOs can define permission boundaries centrally and have the classifier enforce them across agent sessions, said Jha.

That makes the current opt-in period for Enterprise and API users particularly important, as CIOs can establish permission policies before auto mode becomes the default, Jha added.

At the same time, however, CIOs also need to understand the limits of those policies, Jena pointed out.

“Auto mode has three levels of restrictions, with hard-deny rules blocking actions unconditionally, while soft-deny rules can be overridden by a developer’s allow rule. As a result, CIOs should use hard-deny or managed-settings rules for actions that must never be permitted,” he said.

Auto mode changes the security risk profile

Those centralized policy controls, which Anthropic claims should make Claude Code more secure, according to analysts, change the security risk profile of the coding agent rather than eliminate risk.

While the classifier and the policy controls reduce the risk of human failures or everyday risks, it also creates a single point of failure, Jain said.

That means if that single point has a blind spot or an attack manages to bypass it, the agent could still execute harmful actions without human approval or recognition, Jha echoed.

That said, Anthropic is not removing developers’ ability to control how Claude Code operates.

Users who have already set a different default may receive a one-time prompt asking whether they want to move to auto mode, while those who have pinned their existing preference will see no change. The company is also making the transition easier by absorbing the small amount of additional token usage generated by the classifier for Pro, Max, and Team users, with the change effective immediately.

Be careful what you wish for. That’s what Microsoft found out recently when it discovered that AI — rather than making Windows more secure by helping the company close security holes — could help hackers find flaws faster than Microsoft can fix them.

It’s put Microsoft in a serious bind. Should the company devote a potentially massive amount of resources to fix every AI-unearthed security issue as fast as it can? Or should it rush to close the most important ones, and clean up the minor ones later at much less cost?

What the company decides could have tremendous implications. For years, Microsoft has faced criticism from many in the US government because of how often its technologies have been hacked, potentially putting the nation’s security at risk. Some members of Congress have suggested the government curtail contracts with Microsoft until the company proves it can provide more safety. Billions of dollars in federal contracts could be pulled.

To get a better insight into what could happen, let’s look at AI’s newfound prodigious ability to find security holes and bugs.

AI-fueled security mavens Mythos and Project Glasswing

Microsoft’s AI-related security problems are an outgrowth of the launch of Anthropic’s Mythos AI model, designed to handle cybersecurity and biology research. One of its goals is to uncover security vulnerabilities as quickly as possible, so that software companies and cybersecurity companies can fix them before hackers find the holes.

Anthropic’s early tests found that Mythos was spectacularly successful in finding Windows security flaws. “Within 31 minutes, Mythos generated its first proof-of-concept exploit for a Windows kernel vulnerability,” Axios reported in June.

Anthropic noted that its testing of Mythos “reveals a stark fact: AI models have reached a level of coding capability where they can surpass all but the most skilled humans at finding and exploiting software vulnerabilities.

“Mythos Preview has already found thousands of high-severity vulnerabilities, including some in every major operating system and web browser. Given the rate of AI progress, it will not be long before such capabilities proliferate, potentially beyond actors who are committed to deploying them safely. The fallout — for economies, public safety, and national security — could be severe.”

As a result, Anthropic decided to launch a security project it calls Project Glasswing, which it describes as “an urgent attempt to put these capabilities to work for defensive purposes.”

Many companies beyond Anthropic are part of the effort, including Microsoft, Google, Amazon, Nvidia, Apple and others. In theory, it’s a great idea. Find and fix vulnerabilities before hackers can, and everyone is safer. What could go wrong? 

Plenty, as it turns out. Access to Mythos is available to anyone, and Mythos has been finding security holes and bugs far faster than companies — particularly Microsoft — can patch them. As Pro Publica reported in late July, ‘Microsoft is struggling to fix them fast enough.” 

Too little, too late?

For now, Microsoft is only patching the most dangerous bugs and holes. And, according to Pro Public, “internal records indicate that Microsoft plans to eventually address ‘moderate’-severity flaws uncovered by Mythos. The documents made no mention of ‘low’-severity bugs.”

That’s fairly typical of the triage many companies use when deciding how much effort to put into plugging holes. But some experts believe that in the Age of AI, that’s a dangerous way to handle security. 

Vinh Nguyen, former chief AI officer and chief data scientist at the US National Security Agency and now a senior technical adviser to Anthropic and senior fellow for AI at the Council on Foreign Relations, is particularly concerned that the approach is outdated. He told Pro Publica: “The problem now is that you can chain four low-level flaws, and that can equal a high severity. If you’re Microsoft, the current triage strategy may be underpricing risks.” 

It’s already proving difficult for Microsoft to keep up with patching bugs and security holes since Mythos’ launch. July’s Patch Tuesday release fixed the most bugs in Microsoft’s history – 622 of them. Things will only get harder from here. 

The dangers for Microsoft

Microsoft is already in the doghouse for how it’s handled security. A year ago, in one of the worst attacks on Microsoft technologies, Sharepoint was hacked. Tens of thousands of servers were hit, including not just thousands of businesses, but many important government agencies as well.    

The National Institutes of Health (NIH) and the National Nuclear Security Administration (NNSA), which is in charge of the nation’s nuclear security, were among the victims. So were the Department of Homeland Security (DHS), the Cybersecurity and Infrastructure Security Agency, the Transportation Security Administration, Customs and Border Protection, and the Federal Emergency Management Agency, among many others. 

Even before then, some members in Congress were threatening to pull Microsoft contracts if it didn’t improve security. At one point Sens. Eric Schmitt (R-MO) and Ron Wyden (D-OR) sent a threatening letter to the Pentagon requesting it not increase its use of Microsoft products.

“We write with serious concern that the Department of Defense (DoD) is doubling down on a failed strategy of increasing its dependence on Microsoft at a time when Congress and the administration are reviewing concerning cybersecurity lapses that led to a massive hack of senior US officials’ communications,” the letter said, in part.

Nothing happened as a result of the hack or the letter. But that could change in a heartbeat if there’s another big attach. It seems inevitable that hackers will start taking advantage of how quickly Mythos discovers security holes and bugs, especially if Microsoft doesn’t fix them all. 

The next big hack, powered by AI, could prove very dangerous — not just for businesses and governments, but for Microsoft as well.

My current model-selection strategy is embarrassingly simple… and wrong: I choose the most capable (expensive) model because I’m worried the cheaper one might get something wrong. That makes me spend too much, of course, which is why I (like you, perhaps?) continue to struggle with one of the most fundamental challenges in AI: How can I know when a cheaper model is sufficient for a task? Or, really, which model should I use at all?

I asked a friend, Leo Zheng, who leads marketing for Fireworks AI, an AI infrastructure company that runs and improves open-weight models. Surely it was his job to know? His answer surprised me. I thought the answer would come down to models, but it doesn’t. Fireworks, he said, wants to “enable every company to own the continual learning loop within their four walls.” The idea is to abstract away model updates while companies feed the system new signals as customer behavior changes.

That’s when it clicked. I was asking how to automate model choice, but the harder problem is building a feedback loop that tells a company what worked. Model choice then becomes simply one important action the system can take, not the be-all and end-all decision a developer must get right in advance. Arguably, the more important component is integrating enterprise data into that continual learning loop that Zheng describes.

I’ve made that argument before, and I was right. Sort of. Proprietary data doesn’t automatically guarantee AI success. Neither does dumping that data into a retrieval-augmented generation (RAG) system. Your competitors can do the same. The more durable advantage comes from outcome data: traces that connect what the system saw, what it did, and whether the business result was acceptable. Those traces become valuable when a company uses them to improve the system and carry the learning to the next model.

Make the data earn its keep

Public benchmarks are useful, but they only tell you how a model performed on someone else’s test. They can’t tell a bank whether a fraud alert was correct, inform a retailer whether an AI agent resolved a return, or tell a software company whether AI-generated code survived review. Enterprises already have pieces of those answers buried in support transcripts, accepted patches, abandoned carts, etc. The trick is turning that mess into something a system can learn from.

Think about a software company. It already collects product analytics to improve its product. AI can extend that flywheel. Application signals can improve the system powering the product. How? By capturing user intent, engagement, and preferences.

This isn’t foolproof, however. A fast learning loop built on a weak proxy is simply a fast way to make the system worse. For example, suppose Toni in customer support accepts a suggested response. Was it good? Maybe. But what if Toni was rushing (as she is prone to do) and the customer reopened the ticket the next day? It’s not enough to act on inputs alone; a company needs to connect the suggestion to the eventual outcome or risk rewarding the wrong behavior.

Acceptable or ideal outcomes must be defined by the enterprise. A platform like Fireworks can optimize through learning loops, but only if desired outcomes are defined. This changes the cost calculation. After all, the cheapest model by token can be the most expensive model by outcome if people must review, retry, and repair its work. A company can be hemorrhaging money on AI failures while congratulating itself for lowering inference costs.

Open weights get useful

None of this requires open weights, of course. A company can own its evaluations, improve prompts and retrieval, and route among closed APIs and open models without changing a single parameter. For many workloads, that will be enough.

This generally won’t start with reinforcement learning, either. It starts by picking one workflow with an observable outcome and logging the request, model, cost, answer, and what happened next. Do that for a quarter and patterns should emerge. A cheaper model will be good enough for some work, while a proprietary frontier model will earn its price elsewhere. However, in a third category, every model may fail in the same repeatable way, and it’s this last category that makes the serious case for changing the weights.

A few years ago, I argued that open source succeeds when it makes developers’ lives easier, not when it wins an ideological purity contest. The same applies to open weights. The importance of open weights isn’t that they’re a cheaper facsimile of whatever OpenAI and Anthropic are shipping. Instead, they matter because they shift leverage away from the models and toward users with proprietary outcome data. They let enterprises turn what they learn into something they can shape, keep, and carry forward as models change.

Open weights turn model renters into model owners. That’s a very big deal.

For Cursor’s Composer 2, the company started with the open-weight Kimi K2.5 model, then specialized it through continued pretraining and reinforcement learning in environments built to resemble Cursor itself. The interesting thing isn’t that Cursor picked the “right” model, but rather that Cursor can use what it learns from developers to keep making that model more “right” for them. In this case, Fireworks ran the reinforcement learning inference, while Cursor supplied what only Cursor could: its product environment and definition of good. Cursor subsequently built Composer 2.5 on the same Kimi K2.5 checkpoint and says 85% of its compute went into additional training and reinforcement learning.

The base model was the starting point, not the finished product.

This is why open weights matter. They turn the model from a finished product into raw material. The foundation labs have broad knowledge, far more than any normal company could hope to collect. What they don’t know is whether your patch shipped, your fraud was stopped, your customer stayed, or Toni’s response made matters worse. You do.

RAG lets a model consult your data. A learning loop lets the system learn from it. One helps answer today’s question, but the other helps the company answer tomorrow’s question better.

Control must be convenient

Raw weights alone won’t deliver that control. Open source didn’t take over the enterprise because CIOs wanted tarballs. It won because Red Hat, and especially cloud providers, turned theoretical control into everyday convenience. Open-weight AI needs the same translation.

Enterprises want the convenience of a closed API without surrendering control over the intelligence their data creates. That’s the real Fireworks opportunity. Not faster tokens, but a shorter distance between learning something about your customers and making the product smarter because of it. Fireworks CEO Lin Qiao calls this “application-specific intelligence,” and that’s exactly right. Enterprises don’t need to outbuild OpenAI or Anthropic. Even if they wanted to, they couldn’t. Instead, they need to turn what they know about their customers and workflows into intelligence those labs can’t supply.

Fireworks isn’t alone. Together AI also spans training, evaluation, and serving, while Baseten now talks about connecting production feedback to model training. This is becoming a category, not simply a Fireworks feature. Fireworks’ particular bet is that the entire learning loop should belong to the customer.

So, back to the question I was asking Zheng: How do I pick the right model in advance? By now it’s clear: I can’t. Neither can you, and that’s fine, because it’s not really the right question.

Once the enterprise owns the test, it no longer has to trust a vendor’s benchmark or brand. New models must prove themselves against the work that matters to that company, then be promoted, specialized, or discarded accordingly. Model choice becomes a consequence of learning, not an act of faith. In effect, the models audition for the enterprise, not the other way around.

No one can predict whether OpenAI, Anthropic, Google, or the open-weight community will lead six months from now. The sensible response is to stop making that prediction central to the strategy. The enterprise’s own data should decide which model to pick. That’s the bet I find compelling in Fireworks: Make open-weight models as easy to improve as APIs are to call, so enterprises can stop treating their data as context for somebody else’s model and start using it to build intelligence they control.

I came across a LinkedIn post the other day that described “hypegineering,” which the poster explained refers to the moment when AI “marketing becomes more innovative than the technology itself.”

That post came from Ralph Aboujaoude Diaz, the global head of GRC for British consumer services company Haleon. Until last year, Diaz worked in operations cybersecurity for consumer goods giant Philip Morris. 

In his post, Diaz added that “side effects may include believing mediocre tech is revolutionary, confusing hype with progress, buying solutions to problems you don’t have and defending it like your job depends on it.”

As amusing as that might seem, the problem is frighteningly real. The sad truth is that enterprise IT executives are partly — perhaps mostly — to blame for the current hype around AI.

For many decades, senior IT leaders (pre-dating when it was called MIS) were the technology hype-deflators with razor-shape BS detection skills. That level of skepticism was needed. Tech vendors have always exaggerated and left out critical context when they weren’t outright lying about their products. 

Enterprises trusted the IT gate-keepers to ferret out reality from the smoke and mirrors. 

But there was a critical difference back then: senior management (especially CEOs, CFOs and board members) didn’t pay much attention to tech. They wanted the benefits, but they left the details to IT management to sort things out. With senior managers, benign neglect can be an incredibly good thing. 

As much as we might all think that we want our bosses to really care about our efforts, be careful what you wish for. (For Dilbert fans, think of the pointy-haired boss; a boss who has strong beliefs but no understanding of technology is a nightmare.) 

Alas, that is the situation many enterprises find themselves in today. IT management fully understands the level of absurd hype coming from a multitude of AI players. But unlike years and technologies past (RFID? NFC? Biometrics?), deflating hype balloons is easier when it’s comes solely from vendors. When senior managers start spouting this garbage, IT’s hype-deflation ability morphs into contradicting the very people at the top of their own corporate food chain.

That forces IT leaders who want to stay gainfully employed to do a lot of what used to be called lying. “Well, boss, yes these agentic systems have guardrails that will block destruction,” the lie begins, “but we can’t anticipate what any user or attacker might say in a prompt.”

That’s far more corporate acceptable than the actual truth: “Boss, a guardrail that an agent can disregard isn’t a guardrail. It’s more of a mild suggestion.” 

Or IT could say, “Well, yes, boss, autonomous agents could exponentially increase efficiency — as long as you’re OK with the risk that they might send our internal data to the competition.”

For the sake of the company, tech leaders and decision-makers need to reassert themselves and perform serious reality checks on all AI efforts. But this is more fraught than merely contradicting a top boss. IT  could be seen as embarrassing that top boss by pretty much proving that they were either naive or ignorant enough to be conned by the hype. 

Telling them they’re wrong seems nicer than calling them stupid. And yet, someone in IT has to find a politically palatable way to do both. 

For years, AI progress was measured by scale: more data, larger models, and more compute. That formula produced real breakthroughs, but autonomous mobility was one of the first industries to discover its limits.

On the road, AI does not fail quietly. A misread scene, an ambiguous gesture from a pedestrian, or a construction zone that does not match prior examples can create immediate and visible risk. That pressure forced autonomous mobility teams to confront a reality the rest of enterprise AI is now beginning to face: the hardest problem is not access to models. It is reliable ground truth.

As AI moves from pilots into production systems, organizations are discovering that performance depends not only on model capability but on the quality, consistency, and defensibility of the data used to train, evaluate, and improve those systems.

The scaling myth is breaking in production

Early autonomous vehicle development followed a familiar playbook: collect more data, train larger models, and improve performance over time. That approach worked to a point. But real-world driving data does not behave like benchmark data. Road environments are messy and unpredictable. Human behavior is inconsistent. Context changes quickly. Many situations are ambiguous even to trained human observers.

As datasets grew, teams often discovered that they were not simply collecting more signal. They were also collecting more contradictions. Different teams interpreted the same scenes differently. Edge cases accumulated, and ambiguity multiplied. Instead of producing clarity, scale sometimes introduced confusion.

The same pattern is now repeating across enterprise AI. Organizations have access to increasingly capable foundation models, but many AI initiatives still struggle when deployed into production environments. Production exposes weaknesses that benchmarks rarely capture: inconsistent data, ambiguous edge cases, shifting context, and gaps between model behavior and human expectations.

The result is a growing gap between model access and operational reliability. More data is still valuable, but only when it is curated, consistently interpreted, and tied to clear operational definitions. In many production settings, low-quality or inconsistently interpreted data introduces noise faster than models can resolve it.

Real-world AI is multimodal by nature

Autonomous mobility also exposed another reality that is now spreading across AI: real-world intelligence is multimodal. A vehicle does not understand the road through images alone. It must reconcile camera feeds, LiDAR, radar, maps, localization signals, motion history, weather conditions, and human behavior into one coherent interpretation of the scene.

The same requirement is emerging across other high-consequence AI domains. In healthcare, systems may need to connect medical imaging, clinical notes, lab results, and patient history. In agriculture, models may combine satellite imagery, drone footage, soil data, weather patterns, and field observations. In manufacturing and robotics, AI systems increasingly need to reason across video, sensor telemetry, 3D spatial data, maintenance logs, and human instructions.

This raises the standard for ground truth. The question is no longer simply whether an object is correctly labeled in an image or whether a text response is accurate. The harder question is whether the system correctly understands a situation across multiple signals, some of which may be incomplete, noisy, contradictory, or changing over time.

Multimodal AI does not just need aligned datasets. It needs aligned interpretation, and that cannot be engineered into the model after the fact.

Edge cases are not the exception

One of the most important lessons from autonomy is that edge cases are not a small part of the problem. In high-stakes AI, they are the problem.

AI systems often perform well on averages, but real-world systems fail on exceptions. In autonomy, those exceptions may include a pedestrian behaving unpredictably, a construction zone that does not match prior examples, a partially occluded object, unusual road geometry, or an interaction where human intent is unclear. These situations represent a small fraction of total driving events, but they dominate risk.

The same principle applies beyond mobility. In healthcare, it may be a rare presentation of disease. In finance, it may be an unusual transaction pattern. In manufacturing, it may be a combination of operating conditions never previously encountered. In robotics, it may be a physical interaction that looks simple in simulation but behaves differently in the real world.

Organizations often discover that the final increments of reliability require dramatically more effort than the first 90% of performance improvement. This final gap is why many teams find themselves stuck in “hill climbing” — expending enormous effort for increasingly marginal gains.

In multimodal systems, this effect compounds. Each additional sensor stream introduces new edge-case permutations, and reconciling conflicting signals requires expert judgment.

Ground truth must be defensible, not just labeled

Autonomous-vehicle programs that got this right learned that ground truth cannot be treated as static input: a label, a bounding box, or a classification. In the scenarios that mattered, the correct interpretation of a scene had to be reasoned about, reconciled across sensor streams, and defended against alternative readings.

The same standard now applies in other critical settings. A medical annotation may require reconciling imaging findings with clinical context. A financial model’s training data may need to account for regulatory intent, not just transactional patterns. A robotics system operating in an unstructured environment may need ground truth that captures not just what happened, but the chain of causation leading to that moment.

In these cases, ground truth is not simply data that has been labeled. It is data that has been structured, reviewed, and validated through expert judgment. It needs to explain not only what the correct answer is, but why that answer is correct.

This distinction matters because AI systems increasingly operate in environments where being approximately right is not enough. The more consequential the use case, the more important it becomes to create ground truth that is consistent, auditable, and operationally meaningful.

The shift from annotation to judgment

Perhaps the most significant transformation happening in AI today is the evolution of human involvement. The data labeling industry is moving away from simple, crowd-sourced annotation toward more specialized work: scenario design, failure analysis, model evaluation, red teaming, reasoning validation, and edge-case identification.

This work requires people who understand context, ambiguity, intent, and risk. In high-stakes environments, the key questions are not simply whether a system reached the correct answer. Organizations also need to understand why the system behaved the way it did, whether that behavior can be explained, and how to prevent similar failures in the future.

Trust cannot be added through marketing, interface design, or post-hoc explanation alone. It needs to be engineered upstream, embedded throughout the life cycle from data creation to training, evaluation, monitoring, and auditing.

The difference between a model that looks correct and a model that behaves reliably often comes down to whether expert judgment has been systematically incorporated into development.

Software provides scale, and expertise provides judgment. Reliable AI needs both.

What autonomous mobility teaches the rest of AI

Autonomous mobility is often treated as a specialized vertical. It should be treated as an early warning system.

It showed what happens when AI enters environments where the cost of error is high, edge cases dominate risk, multimodal evidence must be interpreted correctly, and opaque decisions are unacceptable. That is exactly where enterprise AI is now heading in healthcare, finance, critical infrastructure, robotics, agriculture, and other high-consequence domains.

The lesson from autonomy is not limited to vehicles. It is that real-world AI must handle ambiguity, edge cases, and multimodal evidence at the same time. Models can generate outputs, but defining what is acceptable remains a human and operational decision.

The next major leap in applied AI will not come from model size alone. It will come from the systems that create, validate, and continuously refine defensible ground truth.

Those systems depend on three things: high-quality data, expert human judgment, and disciplined operational processes. The organizations that recognize this early will have a significant advantage as AI moves from impressive demonstrations to dependable real-world systems.

Autonomous mobility learned this first.

—

New Tech Forum provides a venue for technology leaders—including vendors and other outside contributors—to explore and discuss emerging enterprise technology in unprecedented depth and breadth. The selection is subjective, based on our pick of the technologies we believe to be important and of greatest interest to InfoWorld readers. InfoWorld does not accept marketing collateral for publication and reserves the right to edit all contributed content. Send all inquiries to doug_dineley@foundryco.com.

It used to be that the value and reputation of coding contest winners was very high. That’s no longer the case. Who cares how fast you can produce code that implements an algorithm or solves a problem? Your competition is no longer other people, at least not by themselves, but AI combined with insightful prompting and good taste.

Historically we wrote all our code ourselves and then we tested it ourselves. Yes, there was a period when we outsourced punching our cards to what was essentially a secretarial pool, but we were still writing out the code and doing all the thinking involved in the programming. Over the years we outsourced some of the testing to a QA department, allowed software to guess at what we were typing and complete it, and allowed code-checking software to run in the background as we typed.

Allowing software to help us by guessing what we mean advanced from word completion to line completion to function completion to program generation over a matter of about a decade. Currently a good model running in a good agent harness or IDE can look at your code, suggest improvements, compile and test the improved code, iterate on that, and then come up with something better, faster, more efficient, and more solid.

That shifts the burden from actually programming to reviewing and testing the code. If you don’t even review and test the code, then what you’re doing is vibe coding. Vibe coding doesn’t give you a lot of control over the end product. It is only really efficient at generating quick prototypes and creating technical debt. It’s just not enough structure.

One attempt to add more structure and to control what the model actually generates is specification-driven development (SDD). SDD works, because a spec grounds the model in a single source of truth. However, SDD often comes at the expense of taking too much effort and really being overkill for what you usually need.

Developers are still floundering, wondering how to strike a balance between vibe coding and spec-driven development. One possible approach is behavior-driven development. Another is a variation on test-driven development, sometimes called “facts first.” A fact in this context is an executable invariant tested by a machine. Finally, there’s intent-driven software development (IDSD), which is a new take on the old craft of intent, context, and expectations (ICE).

Meanwhile, we have perhaps a dozen options for IDEs and another dozen options for agent harnesses and many dozens of models that do a good job of generating code. Here I’ll provide a brief tour of six of them: GitHub Copilot, Google Antigravity, JetBrains Air, Kiro, Zed, and Zenflow.

GitHub Copilot

First offered as a plug-in to Visual Studio Code in 2021, GitHub Copilot was one of the first AI plug-ins for coding assistance. Today, you can use GitHub Copilot in a slew of IDEs for code completion, generation, explanation, and debugging. In addition to VS Code, GitHub Copilot is available for Visual Studio, Vim, Neovim, and the JetBrains suite of IDEs. You can also use Copilot on the command line, both through the GitHub Copilot CLI and as an extension to the GitHub CLI. You can also use it directly on the GitHub website.

Currently there are about two dozen supported AI models in GitHub Copilot, including models from OpenAI, Anthropic, and Google. In Visual Studio Code you can add more models than the ones available by default with your Copilot subscription. For example, if you open the Manage Language Models panel in VS Code, you can select models from Foundry Local via AI Toolkit, GitHub Models via AI Toolkit, Microsoft Foundry via AI Toolkit, and, at least in my case, local models via Ollama. Local Ollama models don’t count against your Copilot plan, but they do occupy GPU or unified memory, depending on your hardware configuration.

GitHub Copilot currently has three default operating modes: Plan, Ask, and Agent. Plan mode researches and outlines multi-step plans. Ask mode chats with you without making changes. And Agent mode edits files in your workspace. You may also have custom modes, some of which are supplied by plug-ins; I have 16 of these. You can create your own custom modes if you wish.

GitHub Copilot currently has nine categories of default tools used by agents; each category may have many actual tools listed. Additional tools come from connected Model Context Protocol (MCP) servers and installed plug-ins. My installation currently has 346 tools available, which is nuts. At one point VS Code would warn you if you had too many tools active, but that doesn’t seem to be happening anymore.

AI agents are autonomous helpers that perceive their environment, decide on a course of action, and execute it. They break large tasks into smaller steps, draw on available tools and resources, reason about approach, and adjust their plans on the fly, all while following user directives until the goal is met. While most agentic code edits complete in seconds or minutes, some can run for hours with only a few requests for permission.

In light of that, it’s not surprising that GitHub switched to usage-based billing for GitHub Copilot. Effective June 1, 2026, GitHub tracks your token consumption rather than your Premium Request Units. Code completions and next edit suggestions are exempt from token limits.

Google Antigravity

The original Google Antigravity (from way back in November 2025) was built as an agentic development platform, essentially a version of the VS Code “Code – OSS” IDE that Google developers use internally, which was repurposed for the sorts of repositories that mere mortals work with on normal-sized products. (Google’s internal software mono-repository is so large and complex that ordinary software tools can’t work with it; Googlers use Piper, an internal tool for version control, rather than Git.)

On May 19, 2026, Antigravity became the basis for a new ecosystem. Google Antigravity 2.0 is a standalone desktop application that is an independent, agent-focused surface, evolved from Antigravity’s Agent Manager. The new Google Antigravity IDE is a separate download that looks similar to the previous Antigravity. The Google Antigravity CLI is a way to invoke, monitor, and interact with Antigravity agents from your terminal. In addition, the Google Antigravity SDK is available to build custom agents.

Google Gemini 3.6 Flash is the default model on Google Antigravity. According to Google it’s their strongest agentic and coding model yet, as well as being faster than other frontier models on Antigravity. Gemini 3.6 Flash can be selected at three effort levels: low, medium, and high. Antigravity also supports Gemini 3.5 at low, medium, and high levels; Gemini 3.1 Pro at two effort levels; Anthropic’s Claude Sonnet 4.6 and Opus 4.6 with thinking enabled; and GPT-OSS-120B at medium effort. The general guidance is to use the lowest level of effort that will successfully execute your task to minimize token use.

The “+” dropdown in the prompt box allows you to add context to the prompt. Both Antigravity 2.0 and Antigravity IDE allow you to add media, mentions, and actions. Antigravity 2.0 also allows you to enable the browser from the “+” dropdown; it simply adds the /browser slash command to the prompt.

The Antigravity browser subagent is the rough equivalent of Playwright or the Microsoft Foundry Browser Automation tool. It can click, scroll, type, read console logs, capture the DOM, take screenshots, and record video. The browser subagent also integrates natively with the Chrome DevTools MCP.

The other internal slash commands for both SKUs are /goal, which means to run until the specified task is completely finished, not asking for intermediate input from the user; /grill-me, which means to ask questions back to align on the specific details of the plan before implementing it; and /schedule, which means to run an instruction as a one-time timer in the future or on some recurring schedule (via Scheduled Tasks). Any skills provided by plug-ins or by installed Model Context Protocol servers can also be selected by typing a slash into the prompt box.

Antigravity 2.0 and Antigravity IDE support skills both at the workspace and global levels. A skill requires a SKILL.md file with YAML frontmatter at the top, and supports optional scripts, examples, and resources. The YAML description field is key to allowing an agent to recognize when the skill is relevant. Skills that come with MCPs or plug-ins take no effort to install. Creating custom skills requires you to add a named skill directory and the SKILL.md file.

JetBrains Air

JetBrains Air is an agentic development environment that allows you to delegate coding tasks to AI agents. It supports four agent providers: OpenAI Codex, Anthropic Claude, Google Gemini, and JetBrains’ own Junie. You can provide your own subscriptions or API keys for OpenAI, Anthropic, and Google to Air, or rely on JetBrains’ hosting via a subscription.

When you create a new task for JetBrains Air, it can run in a local workspace, a Git worktree, or an isolated Docker container. Once you have set it up properly, you can also run a new task in the cloud, starting either from your local machine or from a browser.

You can select an agent provider and model at this time. You can switch the agent only when you create a new task; you can change models within an agent at any time. There are four levels of permissions ranging from plan first to full access. For some agents you can select an effort level. You can provide context to the chat from a number of sources, including files and folders, documentation, Git branches, and several others.

To test Air, I gave it the prompt “try to improve on the performance of this pi calculator. Do your work in a new directory called air.” I used Claude Sonnet 4.6 (1M context window) and high effort with Ask permissions, running the task locally. It did a rather good job, although I got further optimizations later on with other environments.

Kiro

Developed by “a small, opinionated team within AWS” and described as an autonomous agent or virtual developer that learns over time while working independently, Kiro is available both as an IDE (based on Code OSS) and as a CLI tool. Kiro CLI (kiro-cli) doesn’t deal with specs at this point, although it does have a planner agent and agent steering. Kiro IDE (kiro) explicitly supports both vibe coding and spec-driven development (SDD). (See screenshot below.)

Kiro SDD generates three markdown files that together comprise the specification: requirements.md, which captures user stories and acceptance criteria in structured EARS (Easy Approach to Requirements Syntax) notation; design.md, which documents technical architecture, sequence diagrams, and implementation considerations; and tasks.md, which provides a detailed implementation plan with discrete, trackable tasks.

You can also import specs from other systems and iterate on your specs. You can even generate specs based on a vibe-coding session. Ideally, you would create a spec for each project feature.

EARS notation captures user stories and follows the pattern:

WHEN [condition/event]
THE SYSTEM SHALL [expected behavior]

This format is clear and testable. Kiro can generate property-based tests (PBT) based on your EARS-formatted requirements. These are more comprehensive than the usual unit tests.

In addition, Kiro SDD can generate three markdown files that together define the steering for the agents. Steering gives Kiro persistent knowledge about your workspace and its conventions.

A product overview file (product.md) defines the purpose, target users, key features, and business objectives of your project. A technology stack file (tech.md) specifies your chosen frameworks, libraries, development tools, and technical constraints. And a project structure file (structure.md) documents your file organization, naming conventions, import patterns, and architectural decisions.

With my free plan, Kiro IDE currently supports two Anthropic Claude models, Sonnet 4.5 and Sonnet 4.0, plus five open-weight models. It can automatically select models if you wish. The documentation also lists OpenAI’s GPT-5.6 Sol, Terra, and Luna, Claude Opus 4.5, 4.6, 4.7, 4.8, and 5.0, and Claude Haiku 4.5, all of which can be activated with a Pro ($20/month) or better plan.

Kiro IDE supports over a dozen context providers for chat requests, including generic context such as #codebase and #spec and specific context such as #code (snippets). It supports skills that follow the open Agent Skills standard, and Model Context Protocol servers that follow the MCP standard.

Zed

Over the last few years my colleague Serdar Yegulalp and I have both reviewed Zed favorably, citing both its raw speed (it was written in Rust) and its integration with a flock of language models, mostly using a “bring your own key” paradigm. You need to subscribe if you want to use a Zed-hosted model.

The team behind Zed previously created the hackable text editor, Atom, and the software platform for building JavaScript-based desktop apps, Electron, both for GitHub. They also built Tree-sitter, an advanced syntax parsing framework that has become the standard mechanism for code analysis and syntax highlighting across modern code editors. While Atom has been discontinued, both Electron and Tree-sitter are still active.

It’s Wednesday so there’s a new release of Zed out; the current version (as I write) is v1.13.2. Recent milestones include support for Anthropic’s Claude Opus 5, OpenAI’s GPT-5.6, llama.cpp as a model provider, and automatic agent context compaction. Some older major features include support for skills, collaboration, and remote development. If you’re interested in particular features, read the Zed documentation or the Zed GitHub Repository.

To test Zed’s agentic capabilities, I pointed it at the C++ π calculation program I’ve been using to test agentic development tools, then used Zed’s new code-review skill to evaluate the program. After thoroughly reviewing Zed’s proposed changes, I gave it the go-ahead to implement and test them. I used the recommended model at the time, Claude Sonnet 4.6.

Zenflow

Zenflow is a free platform that coordinates AI agents to build software using spec-driven development (SDD) workflows. Another term for coordination is orchestration, hence Zenflow is also described as an orchestration layer.

Zenflow was developed by the Zencoder team. Thus Zencoder plug-ins work in Zenflow, and features from Zenflow (such as guided workflows) have been added to Zencoder. The CEO of Zencoder, Andrew Filev, told me that his team of experienced engineers had been using Zenflow for their own product development for over a year when I questioned whether it is ready for production code.

The high-level description of the relationship between Zencoder and Zenflow is that Zenflow is the workflow brain and Zencoder executes the work. You may have noticed some naming confusion: Zencoder is not only the name of the company and of its AI plug-in for IDEs, but it is also the name of its in-house coding agent, which is one of four options for Zenflow (the others are Claude Code, Codex, and Gemini) and one of about nine models available to the Zencoder plug-in.

When you start a Zenflow project, you’re offered a choice of standard workflows: Quick Change, Fix Bug, Spec and Build, or Full SDD Workflow, depending on scope. The wider the scope, the more structure you need in the workflow to keep the implementation from drifting away from the requirements. You can also define your own custom workflows, perhaps to conform to your shop’s standards.

Zenflow can run multiple tasks in parallel in isolated environments. The agents coordinate within workflows without corrupting your codebase.

Zenflow also automates verification of its changes. Every workflow runs automated tests and cross-agent code review. Failed tests trigger automatic fixes. Your code ships only after passing all the verification gates.

Zenflow projects are broken down into tasks, and those are divided into subtasks and chats. Each task runs inside its own isolated Git worktree. You can view the status of all tasks in Kanban boards or stacked list views.

Choosing an AI coding environment

Which AI-supported development environment should you choose? As we’ve seen, there are plenty of good options.

GitHub Copilot is one of the most mature AI coding plug-ins for VS Code, and it supports a wide range of AI models.

Google Antigravity 2.0 is a standalone desktop application that is an independent, agent-focused surface. Antigravity IDE is a separate download that looks and feels similar to the original Antigravity. Antigravity CLI is a way to invoke, monitor, and interact with Antigravity agents from your terminal. These tools are among the best ways to take advantage of Gemini models.

JetBrains Air is a relatively new agentic development environment that allows you to provide your own subscriptions or API keys for OpenAI, Anthropic, and Google models.

Kiro IDE supports both vibe-coding and spec-driven development flows. With my free plan, Kiro IDE currently supports two Claude models.

Zed is about the fastest IDE I have used, and it integrates with a flock of language models, mostly using a “bring your own key” paradigm. Its advanced features include collaborative coding and remote development.

Zenflow supports multiple workflows, from quick changes all the way up to full spec-driven development. You can also define custom workflows. Zenflow works with the Zencoder coding agent as well as Claude Code, Codex, and Gemini.

There is no winner: It’s really a matter of which tool fits your use case and makes you happy. If I could have everything I wanted in one tool, I would take the speed and power of Zed, the spec-driven development capabilities of Kiro and Zenflow, and the agentic capabilities of GitHub Copilot, Google Antigravity 2.0, and JetBrains Air.