Text to pull request,
with a human still in charge.
Human-Gated Engineering Workflow
ADW is the pattern for turning a plain-English request into a scoped software change: classify intent, isolate the workspace, plan the work, implement, test, and open a pull request. The system can move quickly because the review gates are explicit.
The Paradigm Shift
Most teams are in Stage 1. The competitive advantage lives in Stage 3 — where the engineer's job is designing the autonomous system, not operating it.
Copilot
AI autocompletes your code. You still drive everything.
Pair Programmer
AI writes functions on request. You review and integrate.
Autonomous Pipeline
AI plans, builds, tests, ships. You design the system.
Three Layers, One Machine
Click each layer to see what lives inside. The human only touches the top and bottom — intent in, review out. The middle runs itself.
Publishing Workflow
Turns sources, notes, and editorial judgment into reviewable writing decisions without pretending every draft deserves to ship.
Research Lab
Self-improving experiment engine. Runs prompt experiments, scores outputs against rubrics, keeps winners, discards losers.
Operations Dashboard
Real-time monitoring for the entire agent fleet. Pipeline flows, dispatch tracking, tool usage analytics.
Intent Router
A lightweight NL classifier determines intent — feature, bug, question, or reaction. Routes to the right pipeline in milliseconds.
Nimble Flows
Codex and Claude split the work by job: code-heavy planning, implementation, review, verification, synthesis, and taste checks each route to the lane that fits. Each stage is a fresh instance with zero context bleed.
Quality Gates
Composition tokens enforce a strict pipeline: proof-of-life → code review → E2E verification → ship. No skipping steps.
Self-Healing Tests
When tests fail, the builder agent gets the error output and fixes the code. Up to 3 retry cycles.
Metaprompt Templates
Project-specific instructions encoding accumulated judgment. Every correction becomes a permanent template update.
Cron Autonomy
Scheduled tasks run without any trigger. Daily scans, weekly digests, health checks — the system maintains itself.
Telegram Interface
Natural language input from your phone. No IDE, no terminal, no context-switching.
Approval Gates
Review plans before execution. Approve PRs from your phone. Reject and redirect with a single text.
Taste Corrections
Every correction gets encoded into templates so the system learns your preferences permanently.
From Text Message to Shipped PR
Walk through a real task flowing through the system. Click each step or use arrow keys.
A message arrives on Telegram. Plain English, typed from your phone while walking the dog.
Brandon: build: add trending topics sidebar to the newsletter
Every Correction Makes the System Smarter
This is what separates an autonomous pipeline from a chatbot. The system accumulates your taste, preferences, and judgment over time.
Agent executes from templates
Human checks output quality
Human texts specific feedback
Template permanently updated
The agent reads the project's metaprompt template before starting work. This template contains every past correction, style preference, and architectural decision — accumulated over dozens of cycles.
Hybrid model runtime
The pipeline uses two CLI-backed lanes across Codex and Claude, so code changes, review passes, verification, and higher-level reasoning can route to the tool that fits the step.
Inspired by the agentic layer pattern from IndyDevDan and the Stripe engineering team. Built with a Codex and Claude workflow.