The autonomous
workflow

The agentic development workflow: seven phases plus optional Intent, three feedback loops, one PR at the end.

Artifact Chain

What flows between phases

Intent spec

normalized requirement with problem statement, acceptance criteria, and scope boundaries

Markdown

Plan spec

structured implementation plan with file:line targets

Markdown

Build report

changes made, tests written, verification steps

Markdown

Test report

pass/fail results, coverage metrics, fixes applied

JSON + MD

Review verdict

success/fail, issues classified by severity

JSON

Documentation

updated docs, ADRs, changelog entries

Markdown

Commit log

atomic commits, structured PR with evidence

Mixed

Production Health Report

error rates, latency, anomalies, feedback issues

Markdown

At a Glance

Phase summary

Phase	Role	Input	Output	Tools	Loop?
Intent	Requirements Analyst	Raw request	Intent spec with acceptance criteria	ReadSearch	--
Plan	Software Architect	Task description, codebase context	Structured spec with file:line refs	ReadSearch	--
Build	Software Engineer	Approved plan	Code changes, tests, build report	ReadSearchExecuteWrite	--
Test	QA Engineer	Code changes, test suite	Test report (JSON + MD)	ReadSearchExecuteWrite	Test Loop (3x)
Review	Senior Reviewer	Code changes, test results, plan	Review verdict JSON	ReadSearch	Patch Loop (3x)
Document	Technical Writer	Final changes, review summary	Updated docs, architecture notes	ReadSearchWrite	--
Deploy	Release Engineer	All artifacts	Atomic commits, PR	ReadSearchExecute	--
Monitor	Operations	PR/merge data, workflow metadata	Production Health Report	ReadSearchExecute	Issues -> Plan

Self-Healing

Feedback loops

The workflow is not linear. When something fails, it tries to fix itself before asking a human.

Test loop

Test -> Build -> Test

Trigger: test failure

Action:

Analyze all failures
Spawn builder agent to fix
Re-run full test suite

Retry limit: 3 (configurable)

Escalation: surface to human with full context

Review-Patch Loop

Review -> Build -> Test -> Review

Trigger: blocker issues found

Action:

Generate patch context
Build agent applies patches
Re-test, then re-review

Retry limit: 3 (configurable)

Escalation: full context to human reviewer

Monitor -> Intent -> Plan

Monitor -> Intent (optional) -> Plan (new workflow)

Trigger: workflow failures, excessive retries, or threshold violations

Action:

Generate issue from production health data
Refine through Intent phase (recommended for raw signals)
Feed refined intent to Plan phase
Start new workflow run

Limit: N/A (new workflow each time)

Purpose: continuous improvement loop

The Monitor-to-Plan loop is human-mediated by default. Automated feedback requires explicit safeguards. See Monitor Phase for details.

Deep Dives

Explore each phase

Each phase has its own role, methodology, tool permissions, and prompt template. Click any to dig in.

The phases below are reference examples. Your workflow will look different. Use this framework to define phases, gates, and loops that fit your team's workflow and tooling.

Intent is optional at every complexity level. Add it when your input quality varies or when processing monitor-generated issues.

00

Intent (optional)

Requirements Analyst

Clarify first, plan second. Takes any-quality input and produces a well-defined intent specification through codebase research and interactive clarification. Skip when the input is already well-defined.

Read Search Execute Write

01