The model got better. The environment had to catch up.

01 — The model is becoming part of the work

Older models were useful at the edges of a task. They explained code, suggested changes and generated drafts.

The newer ones can remain inside the task longer. They inspect a repository, make a change, run a test, react to the result and continue.

That is a different product category. The value is no longer only in producing an answer. It is in carrying part of the work.

02 — A context window is not memory

Astra’s million-token context gives an agent more room to work before it has to compress its history. It does not make that history durable.

Long tasks produce failed attempts, decisions, constraints and evidence that may matter much later. If those details disappear during compaction, the next context cannot recover them by being smarter.

Persistent memories solve a different problem. The model needs to separate current working state from durable history. More tokens help the agent hold more information. Memory decides what survives.

03 — Prompt engineering is starting to age

OpenAI’s new guidance for Astra recommends shorter instructions, leaner AGENTS.md files and fewer rules written to compensate for weak models.

That feels right. We spent years adding prompts because the models needed constant steering. Some of that scaffolding is now becoming a constraint.

Every workaround for yesterday’s model becomes a rule that tomorrow’s model has to carry.

Prompt engineering will not disappear. It will move closer to product design: clear contracts, useful context and sharp boundaries.

04 — A sandbox is not enough

LocalStack’s acquisition of WonderTwin extends local emulation beyond AWS to services such as GitHub, Stripe and PostHog.

A sandbox answers what a program is allowed to access. An agent environment also needs to answer what happens when the agent believes it accessed something.

A coding agent needs to create a Stripe customer, open a pull request, send a webhook, deploy a function, break a queue and observe the failure. Giving it real credentials is unsafe. Giving it no external systems makes the test meaningless.

The missing layer is a believable fake world.

05 — The world should be disposable

This is more useful than a collection of mocks. The environment needs state. A customer should exist after creation. A webhook should arrive. A trace should show the failure. The agent needs enough evidence to repair the system and try again.

That makes realistic testing possible without spending money, changing production or waiting for another team.

For a small tool that may exist for thirty minutes, a disposable simulated environment can make more sense than a traditional staging setup.

06 — Production still needs a boundary

Simulation does not remove the need for control.

An agent can prove that a change works in a fake environment. Production still needs permissions, approvals, rollback and an audit trail. Akuity’s agentic control-plane work points at this second half of the problem: let agents operate through existing delivery controls instead of handing them raw infrastructure credentials.

The agent should earn access to production by producing evidence, not by asking for a larger secret.

Coda

The model is getting better at execution. The next bottleneck is giving it enough memory, enough reality and enough control to finish the job.

Containers gave software a disposable computer. AI agents will need a disposable world around it.