Intent
Define the outcome, constraints, risk, and the honest meaning of “done.”
Open calls decidedA structured agentic development case study
From vibe coding to a system.
Distilled from building AbsolutePath, this system keeps natural-language speed, then adds what a prompt cannot safely hold: memory, boundaries, independent challenge, and production proof.
4 memory layers
3 independent repos
Role-based agents
Evidence feedback loop
01 / The usual flow
One path. Different roles. Evidence at every handoff.
The owner still begins with plain language. The system converts that intent into a sequence of durable contracts and increasingly expensive proof. Only independent work runs in parallel.
Context
Load current state, code reality, history, known traps, and repository boundaries.
Cold start, not blindContract
Lock one shared contract; split atomic PRDs with OWNS / REFERENCES boundaries.
Dependency DAGBuild
Use one fresh worktree per scope. Implement behind seams; turn criteria into tests.
Build → testChallenge
Inspect independently, try to refute the result, and fix only the affected delta.
Adversarial verifyProve
Run exact-release gates, deploy, exercise the live path, and record honest status.
Production is truthWrite state, evidence, decisions, and lessons back into memory
built on branch→merged→deployed→
production-verified
02 / Repository + memory architecture
Three boxes. Four kinds of truth.
The umbrella sees the system. PRD decides what. App records how. The bridge tells us what is real.
03 / Agent orchestration
One manager. Bounded workers.
Workers build locally. Independent reviewers challenge the frozen candidate. The manager owns outward authority.
Small changeDirect worktree pipeline.
Architectural changePRD before code.
Production actionManager or owner only.
04 / Failure → safeguard
Every guard has a scar.
Four recurring failure patterns, converted into simple operating rules.
FailureMapped a stale checkout
SafeguardFetch origin/main + read the live bridge first
FailureParallel branches collided in shared memory and numbering
SafeguardReserve scarce IDs · one writer · serial integration gates
FailureIn-memory tests passed; real database constraints could fail
SafeguardTest schema changes against the real adapter before merge
Failure“Green” badges hid skipped builds and non-executed CI
SafeguardRead execution evidence · prove the serving artifact moved
05 / The 40-hour signal
Coding was the small part.
Preliminary finding: safety stayed high, but assurance was repeatedly expressed in natural language.
41 hours of assurance, not 41 hours of code.
Keep the safety. Automate the deterministic repetition.
Next evolution · deliberately only a preview
Proposed, not adopted
01Structured evidenceFacts recorded once
02Content-aware gatesRerun only what changed
03Delta reviewReopen affected lenses
04Agent judgmentReserved for semantic risk
06 / A reusable starting point
Start with four artifacts.
You do not need the full system on day one.
01 / MEMORYSTATE.mdWhat is true now + next action
02 / CONTRACTOWNS / REFERENCESScope and acceptance criteria
03 / PROOFTEST → REVIEW → PRODA visible status ladder
04 / LEARNINGDECISIONS + ISSUESThe next agent starts smarter
07 / Bring it into your project
Give your coding agent this prompt.
It confirms the project, estimates the adoption, helps you choose a scope, waits for approval, then installs and teaches the workflow.
01 / PASTEOpen your workspacePaste the prompt where you normally develop. The current folder is only a candidate.
02 / CONFIRM + ESTIMATEAgent asks, then assessesOne agent runs one bounded read-only pass, then explains effort and likely agent usage.
03 / APPROVE + LEARNYou stay in controlOnly after approval does it install, validate, and guide a small pilot task.
Preview the complete workflow setup prompt
When you use this, the agent will:Ask → Assess read-only → Estimate → Choose scope → Approve → Install → Guide a pilot.