GPT-5.6's Limited Preview Is the Moment the Agent Stack Snapped Together

Start with the artifact. OpenAI shipped GPT-5.6 as a limited preview, and The Sequence's recap leans on that word "precisely" when it makes the distinction. The framing is doing real work.

When a model only answers, the risk surface is bounded by the text it emits. A bad answer is a bad answer. You read it, you discard it, the blast radius ends at your screen. There is no operational reason to gate access slowly, because nothing the model does propagates beyond the conversation.

When a model acts, the risk surface is the world the model touches. Code it writes and runs. Tickets it closes. Purchases it makes. The blast radius is now whatever permissions you handed it. A staged preview is the rational response to that change: you let a small set of operators probe the failure modes before the capability reaches everyone holding live credentials.

So the release mechanics are not a marketing choice. They are a confession. The recap's line that AI is learning "to act" rather than "to answer" is not a slogan, it is the reason the rollout looks the way it does. This maps cleanly onto what we have called the Autonomy Spectrum: deployments run from copilot to full autonomy, and most failures come from deploying at the wrong point on it. A limited preview is OpenAI declining to pick that point for you all at once.

The recap names four directions the industry marched from: better models, richer environments, more autonomous agents, and harder evaluations. Read that through a Wardley Mapping lens and the week stops looking like a coincidence.

Each of those four is a value chain with its own evolution axis, from genesis to commodity. For years they sat at different points. Models were the only mature column; the rest were custom, artisanal, research-grade. That asymmetry is why early agents felt like demos: a strong model bolted to a brittle environment, evaluated by vibes.

What "snapped together" describes is several of those columns reaching the product stage at once. The same week as GPT-5.6's preview, the agent-orchestration tooling that power users actually touch kept shipping ordinary product hygiene. CrewAI's 1.15.1 release added things like initializing Git repositories for generated projects and requiring explicit project definitions. None of that is glamorous. That is the point. Boring, opinionated defaults are what a category looks like when it leaves genesis behind.

The environment thread is moving too. The Sequence's companion piece on self-driving labs describes connecting AI to automated experimental hardware so that "the results of each experiment influence what the system does next," a lab that is "learning while it works" rather than running a prewritten queue. That is the environment column maturing in the most literal sense: the agent's world is now instrumented enough to close the loop without a human transferring the samples.

When multiple columns commoditize together, the bottleneck moves. It moves to whatever is still custom. This week, the thing still custom is evaluation, which is why the recap puts "the future of evaluation" in its own title.

Our standing position on this title is the Harness Hypothesis: the value in AI is not in the model, it is in the harness that connects the model to the world. A week where the headline model ships as a gated preview while the orchestration layer keeps shipping product hygiene is about as clean a confirmation as you get.

Consider what GPT-5.6 changes for a reader who does not build frameworks. Almost nothing, directly. You will not feel a smarter model the way you feel a faster laptop. What you will feel is whether your harness can hold a more capable model without spilling.

The harness is everything between the model and the action: the permission system, the tool definitions, the approval gates, the logging. CrewAI requiring explicit project definitions is a harness decision, not a model decision. It narrows what the agent can assume on its own. That is the layer that determines whether a more autonomous GPT-5.6 is an upgrade or an incident.

This is also where Commoditize Your Complement explains the strategy. The model vendors have every incentive to keep pushing capability into the model and letting the harness layer fragment, because a commoditized harness makes their model the scarce input. The harness builders have the opposite incentive. The reader sits in the middle, and the practical advice is unsentimental: when a more capable model arrives, audit the harness before you celebrate the model. The capability is OpenAI's. The control is yours.

The recap's title flags "the future of evaluation" alongside models and games. That ordering is deliberate. Evaluation is the column that has not commoditized, and it is the one that now sets the floor for everyone else.

Here is the mechanism. As long as agents merely answered, evaluation was tractable: compare output to expected output, score it, move on. Once agents act, the question becomes whether the sequence of actions was correct, safe, and reversible, not just whether the final answer looked right. That is a categorically harder grading problem, and the industry knows it, which is why "harder evaluations" is one of the four converging threads rather than an afterthought.

This is the Capability vs. Controllability Frontier in plain terms: more capable models are harder to control, and the frontier forces an explicit trade-off. A harder evaluation regime is the market's attempt to price that trade-off. When the bar rises from "did it produce a good answer" to "can we prove it acted correctly," every existing deployment gets re-graded against the new bar whether or not it opted in.

For the power user, the recalibration is concrete:

• The agent that passed your informal spot checks last quarter may not pass an action-level audit this quarter.
• "It usually works" is no longer a defensible standard when the model is doing things, not saying things.
• The evaluation you trust has to grade the path, not just the destination.

None of this requires new tooling you do not have. It requires applying the harder standard the industry just adopted to the agents you are already running.

There is a tempting way to read "AI is learning to act," which is to hand it the wheel and grade the wreckage later. The better reading came from a quote The Sequence's neighbor in the feed surfaced this week.

The developer Jon Udell, quoted by Simon Willison, pushed back on the phrase "human in the loop" because "it cedes authority to the machines." His reframe: "It's our loop, we work the same way we always have, now we recruit agents to join the team." An agent-assisted process, he argues, "need not be a black box that takes in prompts and emits features."

That is the correct governance posture for a converged stack, and it is more than a vibe. It is a direct answer to the Shadow Agent Problem: agents installed by individuals without oversight carry the same risk as shadow IT, but with broader system access. The defense is not to ban agents. It is to insist they join a loop you own, with the inspection points intact.

In practice that means the harder evaluation bar and the "our loop" framing are the same advice from two directions. Keep the review gate. Keep the action log. Keep the human as the one who chooses the next experiment, the way the self-driving-labs piece notes a human scientist traditionally "decides what to test" and "chooses the next experiment." The convergence does not retire that role. It raises the stakes of doing it badly.

The reframe also kills the worst failure mode of capable agents: the unreviewable output. Udell's title is literal about it, a doctor's-office joke about agents creating PRs nobody can review. The fix is not a smarter model. It is a loop that refuses to accept work it cannot inspect.

Strip away the framing and here is the operating reality. The model layer got more capable and shipped cautiously. The orchestration layer kept hardening its defaults. The environment layer started closing its own loops. The evaluation bar moved up to grade actions, not answers.

For someone configuring agents rather than building them, that produces a short, specific to-do list.

• Audit the harness before chasing the model. A more capable model amplifies whatever your permission and approval setup already allows. If you would not trust your current gates with a more autonomous agent, the new model is a liability, not an upgrade.
• Re-grade your live agents on action correctness. Stop scoring final answers. Score whether the sequence of actions was safe and reversible. The industry just adopted that bar; apply it to what you run.
• Find your shadow agents. The convergence makes capable agents cheap to spin up. That is precisely when ungoverned ones proliferate. Inventory them before the next preview becomes general availability.
• Keep the loop yours. Treat agents as recruits to a process you control, not as black boxes you query and hope. Every action an agent takes should land somewhere you can inspect.

The convergence story is real, and the recap is right that the threads snapped together into something legible. But "legible" is the operative word, not "finished." A stack that is finally readable is a stack you can finally govern. The release that prompted all of this is still a limited preview. The right response to a limited preview is a limited, deliberate expansion of trust, graded the whole way.