Agentic TDD Harness
Sole engineer. Designed, built, and evaluated the pipeline.
Timeframe
[Pending]
Timeframe: build date, run length, first-commit date.
Stack
Claude Code · parallel CLI agents · always-on hooks
Predates the playbook. The lessons here became it.
Loop engineering, before the name
[Pending]
Section held for the harness first-commit date.
The claim ships with a date or not at all.
Problem.
Agents write plausible code faster than one human can verify it.
Plausible is not verified.
Speed without verification is just risk arriving later.
I wanted ticket in, verified diff out. No babysitting.
[Pending]
Ticket source: the codebase the 101 tickets ran against.
Constraints.
One engineer. No second pair of eyes on demand.
Agents do not judge their own work. Every gate is external.
PRs cap at 100 authored lines. Review speed is the budget.
No code beyond test scope. The tests define the diff.
Approach.
- 1
Draft many plans in parallel. 5 verifiers vote on every one.
- 2
Compete, then commit. Multiple plans race; one winner absorbs expert fixes.
- 3
Build on a leash. RED to GREEN to VERIFY; 2 attempts, then stop.
- 4
Review in force. 9 parallel reviewers, then 14 pre-ship agents.
- 5
Ship at zero findings. Auto-fix loops back to BUILD, 20 times max.
Stages loop until checks pass; hard caps stop runaways.
Context is King. Never Trust, Enforce. Validate Always.
Artifact.
The pipeline
MULTI-PLAN parallel CLI sessions
analyzer + test-scout waves
5-verifier consensus: 4/5, three rounds
SELECT multiple plans compete
expert corrections folded in
optimizers iterate, 3 rounds max
losers discarded
BUILD /tdd: RED → stub-immunity → GREEN
→ VERIFY → SIMPLIFY → COMMIT
2-attempt cap; no code beyond test scope
MULTI-REVIEW 9 parallel reviewers
auto-fix in place, then re-review
20 cycles max
PRE-SHIP 14 agents; expert critics; auto-fix
gate: 0 findings at every severity
auto-fix loops back to BUILD, 20 times max
E2E → SHIP
POST-SHIP /clear-pr-findings until APPROVED
adaptive learning from human reviewsAlways-on guardrails: pre-write, pre-bash, and post-write hooks;
no-stub enforcement; wave gates; spec allowlists.
93/101
tickets completed to a verified PR
0
findings allowed at ship, at every severity
20x
hard cap on auto-fix loops
100
authored-line cap per PR
[Pending]
Example diff and PR artifact.
Results.
93/101
Tickets in, verified PRs out. One pipeline, end to end.
8
Failed during harness build-out; most served as test PRs.
[Pending]
Definition of verified PR: gates and review criteria.
Per-ticket breakdown of the 8 failures.
Learnings.
The lessons became the playbook:
small slices, external gates, hard stops.
Verification is the product. The code is a byproduct.
[Pending]
What I would change: gated on the failure analysis of the 8 misses.