My OS

08-execution-loops.md

Execution Loops

Execution starts only after a project is approved or explicitly marked as an experiment.

The ELT critiques plans. Workspace teams execute plans.

Core Loop

Arijit + Athena approve plan
  -> Athena recruits execution roles
  -> PM decomposes tasks
  -> Architect designs if needed
  -> Developer / Researcher / Writer / Designer executes
  -> QA or reviewer verifies
  -> Tech Lead or Athena signs off
  -> status and decision notes are updated

Single Task Loop

Use for independent work.

Athena assigns task
  -> role executes
  -> role writes context-cache result
  -> verifier reviews against acceptance criteria
  -> pass: mark done
  -> fail: return with evidence

Max loops: 3. After 3 failed attempts, Athena must intervene. The likely problem is unclear scope, wrong role, missing context, or bad architecture.

Dev + QA Loop

For code work, Dev and QA are paired.

PM writes task with acceptance criteria
  -> Developer implements with tests
  -> Developer writes handoff summary
  -> QA runs tests and checks behavior
  -> QA writes pass/fail report with evidence
  -> Tech Lead reviews if risk is material
  -> Athena reports done only after verification

QA reports; QA does not silently fix implementation work unless explicitly recruited as Dev.

Parallel Execution

Parallelize only when tasks do not collide on the same files, assumptions, or decision dependencies.

Good candidates:

  • Independent research briefs.
  • Separate UI screens.
  • Separate source collectors.
  • Separate QA passes.
  • Content variants with shared brief.

Bad candidates:

  • Shared core architecture.
  • Same file or schema changes.
  • Unclear acceptance criteria.
  • Work that depends on one unresolved decision.

Failure Matrix

Failure Handler Escalation
Test failure QA -> Developer Tech Lead after repeated failure
SOP violation QA or Tech Lead -> Developer Athena if repeated
Architecture mismatch Tech Lead -> Architect ELT CTO if strategic
Missing evidence Reviewer -> role Athena if pattern repeats
Context exhaustion Role writes summary and restarts Athena
Spec ambiguity PM clarifies Arijit if product decision
Cost/scope growth Athena pauses execution Arijit

Status Report Format

Athena reports work like this:

Project:
Phase:
Health:

Done:
- ...

In progress:
- ...

Blocked:
- ...

Decision needed:
- ...

Next:
- ...

Artifact:
- ...

Handoff Budgets

Keep handoffs short.

Handoff Max Length
Task brief 2,000 words
Implementation result 1,500 words
QA report 1,500 words
Architecture decision 2,000 words
Research brief 2,000 words
Status update 500 words

Done Means Verified

Athena should not say a deliverable is done because an agent claimed it is done.

Done means:

  • Acceptance criteria met.
  • Evidence exists.
  • Relevant tests/checks passed.
  • Artifact path is known.
  • Remaining risks are named.