Fleet 2.9.1 is out.See what's new →
FleetFleet
Fleet Lessons

Your agents should stop
relearning how you work.

Fleet Lessons turns corrections your existing AI harness already understands into governed, versioned policy. You decide what sticks, where it applies, and when it changes.

Use the harness you already have through MCP or REST. Fleet runs no refinement model and requires no Fleet chat.

Observed pattern

Lesson proposal

Awaiting approval
“Always run the race detector before merging authentication changes.”
Proposed by
Workload via MCP
Evidence
3 independent tasks
Scope
Project
Target
House rule

Exact effect preview

add_requirement · auth.merge_checks · race

Reject
Review exact preview

Your harness learns. Fleet makes the learning durable.

The harness already has the conversation, model, and task context. It recognizes the correction and submits a complete proposal. Fleet takes over at the governance boundary.

Your harness understands

Use the model already in the room

The harness sees the correction, compares relevant candidates, and writes the complete Lesson proposal with the model already serving the task.

Fleet governs

Turn meaning into controlled change

Fleet validates identity, scope, evidence, lifecycle, target authority, revisions, and the exact change that would be applied.

People decide

Keep authority human

An authorized person reviews the exact preview. Workloads can propose and apply approved changes, but they can never approve them.

From correction to policy, without a second AI system

Your harness supplies the intelligence. Fleet supplies the rails.

01

Correct

Tell your existing harness how you want the work done.

02

Propose

The harness submits a strict, complete Lesson through MCP or REST.

03

Review

Fleet checks authority, scope, evidence, conflicts, and the exact target preview.

04

Approve

A person approves or rejects that exact revision. Servers cannot borrow human authority.

05

Apply

Fleet applies direct targets or governs harness adoption, records provenance, detects drift, and can reverse it later.

Scoped by design

One correction, only where it belongs

A preference should not accidentally become company policy. Every Lesson names an authorized scope, and broader promotion is a separate human decision.

Workloads receive organization scope, plus synchronized task identity such as project, team, role, and agent type. They never invent a personal scope.
  1. 1

    Personal

    Private policy for one person.

  2. 2

    Agent type

    Shared behavior for a stable kind of agent.

  3. 3

    Team or role

    Policy for one operating group.

  4. 4

    Project

    Repository-specific policy.

  5. 5

    Organization

    Company-wide policy, explicitly promoted.

Full-scope learning, not a notes folder

A Lesson governs the real artifact that controls behavior. Fleet supports the complete target model instead of trimming every correction down to a personal preference.

Personal policy

Private instructions composed before inference.

House rules

Managed requirements and prohibitions scoped across the organization.

Agent prompts

Versioned changes to managed prompt sections.

Workflows

Exact, validated patches to saved workflow definitions.

Repository prompts

Governed changes through an approved pull request.

Skills

Tenant-owned skill changes with immutable source and adoption proof.

No Fleet model

Keep the intelligence where it already is

Fleet does not discover, refine, embed, or semantically deduplicate Lessons. Your harness does the semantic work with the model you already chose.

The same governed lifecycle over MCP and REST.

No Fleet Chat dependency.

No transcript or chain-of-thought field in the contract.

Optional evidence excerpts are bounded and controlled separately.

Automation can carry policy. It cannot grant itself authority.

Fleet recognizes exactly two identities, with a hard boundary between automation and human decisions.

Human

Interactive identity

OAuth identifies the person using the dashboard, CLI, or MCP. Personal scope and approval stay attached to that person.

Workload

Server identity

Fleet-issued client credentials identify servers and unattended harnesses. They can propose and apply approved changes.

Never delegated

No borrowed authority

A server cannot inherit a person’s scopes. Approval and promotion always return to an authorized human.

Tenant master switch

Per-person opt-out

Observed-pattern control

Evidence excerpts off by default

Every Lesson has a lifecycle—and an undo button

Learning should compound without turning into invisible, irreversible policy.

Exact previews

Approval binds to one revision and the exact rendered target—not a vague summary of intent.

Versioned history

Optimistic revisions and immutable provenance prevent silent rebasing or invisible mutation.

Drift detection

Fleet can distinguish an adopted Lesson from a target that changed afterward.

Explicit reversal

Reject, supersede, retire, or revert without pretending the old policy never existed.

Fleet Lessons, in brief

Is Fleet reading and analyzing every chat?+

No. The harness already serving the task recognizes and refines the correction, then submits a strict Lesson proposal. Fleet has no transcript or chain-of-thought field. Optional evidence excerpts are separately controlled and bounded.

Do I need Fleet Chat?+

No. Lessons is designed for the harness you already use. Fleet provides the governance and lifecycle layer through MCP and REST.

Does Lessons work with Codex, Claude Code, and OpenCode?+

Yes. The hosted Lessons API can be used by Codex, Claude Code, OpenCode, another MCP client, or a direct REST client. Fleet’s coding-agent runner uses Claude Code or OpenCode, including models those harnesses support such as Claude, Codex, and Grok.

Can a server approve a Lesson?+

No. Fleet recognizes human and workload identities. Workloads can propose and apply already-approved changes, but approval and promotion require an authorized human.

What can a Lesson change?+

Personal policy, house rules, managed agent prompts, saved workflows, repository prompts, and tenant-owned skills. Each target has its own validation, authority, preview, application, drift, and reversal contract.

Can we turn learning off?+

Yes. Administrators have a tenant master switch and separate controls for observed patterns, personal Lessons, and evidence excerpts. Each person can narrow those settings further for their own identity.

What happens when a Lesson is wrong?+

It can be revised or rejected before approval, superseded or retired later, and reverted after adoption. Fleet retains the lifecycle and provenance so the reversal is explicit and auditable.

Let corrections compound.

Keep the harness your team already uses. Add the governance layer that turns what it learns into policy you can review, apply, and reverse.