#agents #autonomy #framework

Our operating model, in autonomy levels

Asif Foysal Meem3 July 20266 min readViews

Read as

↳working implementation · tests

Build treatmentUpdated 14 Jul 2026

Agent operations · Earned autonomy

What you will leave with

Classify tasks, write ownership contracts, and run safe parallel work without fleet cosplay.

For engineers turning hand-coordinated agents into isolated, evidence-producing lanes.~3 min read

PrerequisitesGit worktrees or ephemeral workspacesCI with named checksSeveral genuinely independent tasks

01 · System view

What you will build

A manager lane for decisions and worker lanes for bounded execution.

01ClassifyScore risk, agency, and orchestration separately.
02DecomposeFind slices with disjoint files and outcomes.
03ContractEncode ownership, exclusions, proof, and stop conditions.
04DispatchGive every lane isolated state and minimum authority.
05AggregateReview evidence packets and reconcile shared work once.

02 · Implementation

Build it in sequence

Audit actual tasks, not platform aspirations

List what agents do today and place each task on the lowest level that describes its real behaviour. Multiple agents that a person sequences manually are not managed-by-exception.

Answer the three defensibility questions

For every task above supervised delegation, name detection time, rollback, and evidence. Missing answers cap the task's autonomy until a control closes the gap.

autonomy/task-profile.ymlyaml

task: post-deploy-canary
agency: goal-driven
orchestration: parallel-lane
detect_wrong: required assertions against staging
rollback: revert one pull request
proves_right: evidence packet bound to commit and deployment
human_decides: visual taste and production promotion

Decompose for structural independence

Choose lanes with disjoint files, state, and acceptance criteria. Keep shared files with the manager or queue dependent work; do not solve a bad decomposition with retries.

Write the lane contract before dispatch

State owned paths, forbidden paths, required evidence, allowed tools, and stop-and-report conditions. A lane that discovers cross-boundary work must escalate, not silently widen scope.

Aggregate evidence, not transcripts

Require commit, checks, artifact, observations, and recommendation. The manager reconciles dependencies and shared changes after lanes finish rather than refereeing live edits.

03 · Review

Definition of done

Agency and orchestration are scored separately.
Every parallel lane owns disjoint files and state.
Cross-boundary work stops and reports.
Evidence identifies the exact commit and artifact.
Rollback and human decisions are explicit.

04 · Failure modes

What breaks and what it means

Workers collide after other lanes merge.

Ownership or dependency boundaries were not genuinely disjoint; redesign the slices.

The manager becomes a log router.

Replace transcripts with compact evidence packets and decision requests.

Public discussion

Questions, corrections, and useful disagreement

Moderated on GitHub ↗

Sign in with GitHub to join the conversation. Comments are public. Article views are anonymous and counted once per browser, per article, each day.