AGENT ENVIRONMENT •
Framework released
The Agent Environment Framework · Public release

The thinking lives in the environment, not the chat.

A field-tested operating framework for directing fleets of AI agents: persistent state, real tools, explicit permissions, written standards and independent review. Not a product launch. A method for making agents useful inside real companies.

Inspect the evidence room ↗

The framework

An AI session can answer. An agent environment can operate. Give sessions a shared filesystem, explicit permissions, real tools and a review protocol—and the work no longer disappears when a chat ends.

01

Persistent state

LEDGER.md · STATUS.md · HANDOVER.md

Keep decisions, progress and handovers in files that survive a session ending. Dated records give the next session a starting point and a trail it can check.

02

Forced metacognition

bookends · in writing · before start, before done

Ask “how would this claim be wrong?” before starting and again before declaring done — in writing. Make the assumptions, evidence and unresolved limits visible to the next reviewer.

03

Independent review

counters · frozen candidates · duplicate protection

Freeze the work being reviewed, give another session the evidence, and record its findings. Keep review counters and receipts so the next session can see which checks have already run.

04

Different models

one authors · a second reviews · a third reviews independently

Have one model author the work, a second review it and a third review it independently. Record disagreements and resolve them against the evidence before release.

05

Tools with boundaries

email · web search · files · company systems

Give agents the tools the work requires. Define what each seat may read, change and send. Keep external actions behind explicit authority and record the result for the next session.

The operating loop

A company-shaped way to run agents.

  1. 01 / SEAT

    Define the role.

    Give the session a named responsibility, its rules and the current state.

  2. 02 / EQUIP

    Grant the tools.

    Set permissions for the files, systems and external actions the role requires.

  3. 03 / EXECUTE

    Work in writing.

    Put decisions, evidence and artefacts into the shared environment.

  4. 04 / REVIEW

    Challenge the work.

    Use a fresh session or another model to find what the author missed.

  5. 05 / HAND OVER

    Carry it forward.

    Record what changed, what remains and what the next seat must know.

A good firm is smarter than any employee in it. Sessions are the employees. The filesystem is the institution.

Built by operating it

Jesse runs his companies with a fleet of AI agents he directs. This framework came from daily use inside those companies—built by an operator with no coding background and no formal research.

12 months

of daily use inside live companies.

Grace

The environments built Grace — customer-facing AI deployed in legal and finance.

1,000s

of enquiries handled by deployed agents.

Companies the fleet is responsible for
No Lead Lost on LinkedIn →
From Jesse · Framework author · October 2026

I didn’t start with a thesis.

I was deploying agents inside my own business. I had no option but to make that investment pay. This method grew out of the operating pressure to make it work, day after day.

For the last several months, I’ve been bullish: agents need an environment that carries the state, the standards and the work forward. It’s the only way I would deploy agents into my business.

As this idea gains traction, I want the actual operating method in the open. The useful conversation is no longer which model is smartest. It is how we build the institution around many models.

For me, this was an operating result before it was a thesis.

— Jesse

For press, builders & operators

The next leap is not a smarter chat. It is a better institution.

Discuss the framework

No product to buy. No waitlist. Just the method, released by the operator who runs it.