go-junie

Agentic Coding Reference

Agentic coding that amplifies an engineer's judgment instead of replacing it.

Ship in days what would otherwise die in triage: work worth trying but not worth weeks, built and tested against real users instead of shelved. The machinery that makes that repeatable — durable specs and nested feedback loops that keep every agent, session, and person pointed the same way — is the substance underneath.

The shape, in one minute. A file-based pipeline of nine one-job specialist agents builds one vertical slice at a time. Each appends a schema-validated record to a shared log, a coordinator routes from it, and four reviewers plus a change-grader gate every change — nothing auto-merges. The work runs through four nested feedback loops, from the inner TDD cycle out to whole-codebase review, so drift is caught before it compounds. Durable specs — PRD, system design, ADRs, ubiquitous language — are the shared memory every agent, session, and person reads and writes. One CLAUDE.md carries it across four agent tools; /materialize and /harvest adopt it in your project and feed improvements back.

AI coding agents face the same two challenges human engineers always have: keeping long-term memory across sessions, and running multi-scale feedback loops that catch drift before it compounds. The difference is degree, not kind — a human forgets between Friday and Monday; an agent forgets between one message and the next. Within days, not years, an agentic project that skips the disciplines that compensate starts drifting: terms picked inconsistently session-to-session, settled decisions re-litigated, this week's architecture contradicting last week's.

The fix is to treat the disciplines human teams already built for these problems as the memory and feedback substrate — documentation standards, DDD, TDD, ADRs, ubiquitous language, and XP-style nested loops. Every agent, every session, and every person on the codebase reads and writes the same durable specs, so all stay pointed the same direction. A file-based specialist pipeline of nine one-job agents operates it, building one vertical slice at a time, and a single rules file (CLAUDE.md) carries it across Claude Code, Copilot CLI, OpenCode, and Junie CLI.

Two working reference implementations (Go, Spring Boot), portable skills, and enforceable documentation standards demonstrate the pattern; a bidirectional /materialize + /harvest loop adopts it in your own project and feeds improvements back.

It is for anyone running an agentic coding workflow over more than a few sessions: a solo developer driving an agent team past what fits in one conversation, a team where each developer drives their own agent team on a shared codebase, or a human-only team that wants the same discipline against the slower drift humans face. The failure modes are the same; only the speed differs.

The architecture, principles docs, and reference implementations are stable and in active use. The specialist pipeline machinery (JSONL contract, four-reviewer fan-out, capability progression) is operational, though its cost-effectiveness is still being measured (with Harness Stats) and will be revised as evidence accumulates. Treat the disciplines as the validated core and the pipeline machinery as one reference implementation of the shape the harness can take.

→ Deep dive: agentic-harness.md covers the loop model and handoff contract. specialist-agent-workflow.md covers the full architecture and migration playbook. the document-writing skill covers the writing rules that keep agents from guessing.

The sections move from how it works to trying it to reference — read top-down, or jump to what you need.

The Force Multiplier and Your Part in It

An agent multiplies whatever it is pointed at. Judgment is the scarce input — you supply the design and the standards; the harness supplies the memory, discipline, and execution that amplify them. The disciplines that keep the multiplier raising quality rather than noise — TDD, DDD, owned specs, ADRs — matter more at this speed, not less.

How the work divides: