Agentic Engineering

How I structure codebases, bots, skills, and review loops so agents can do real engineering work without turning the repo into a lottery machine.

The Evolution

I started building these before cross-app orchestration tools were really a thing. Davy Jones Intern was built right when the Claude Agent SDK first came out. Each one fed into the next.

Davy Jones Intern

Built when the Claude Agent SDK first launched. Slack bot for our team at ZKP2P that handles PRs, CI, docs, Linear tickets, onboarding. Runs on a home server with a codebase knowledge base injected into every interaction.

Barbossa

Built before tools like Devin shipped. Six agents (Engineer, Tech Lead, Discovery, PM, Auditor, Spec Generator) that coordinate through GitHub and Linear. Pick tasks, implement, review, merge autonomously.

Oneshot CLI

Give it a repo and a task. Plans with Claude, executes with Codex, reviews the diff, opens a PR. I dispatch across multiple repos in parallel over SSH.

Oneshot Bot

Slack and Linear control plane wrapping Oneshot CLI. My daily harness at ZKP2P. It dispatches coding work, tracks PRs, routes status back into Slack, and turns the product engineering loop into something I can operate from the team channel.

Custom Workflows

I structure every repo for agents, not for me coding manually. CLAUDE.md files, rules, and custom skills that let agents operate autonomously. The goal is that an agent can debug a stuck deposit by querying our indexer API or check what the arb bot is doing via Firestore without me being in the loop. Two examples of many.

zkp2p-clients

Main ZKP2P client codebase

  • Debug skill that queries the indexer GraphQL and curator REST API to investigate stuck deposits, expired intents, taker tiers
  • Release skill for branch promotion from main through staging, preprod, to prod
  • SDK publishing and browser extension build skills

galleonlabs-zkp2p

UsdcToFiat, Peerlytics, Delegate

  • Bot-state skill reads Firestore to check Delegate's rate decisions and intent activity
  • Logs skill tails Firebase and GCloud for the arb bot, notifications, analytics sync
  • Rules for rate management strategy, auto-generated files, deployment conventions

Paperclip

Paperclip is not my project. I use it as the orchestration layer for autonomous management of repos like Peerlytics and USDCtoFiat. Those agents run on heartbeats, triage their own backlog, ship PRs via Oneshot CLI, review each other's work, and deploy to production. I check in when I want to, not because the system needs constant steering.

Captain4hr

Triages, prioritizes, creates issues

Bosun30min

Picks issues, dispatches Oneshot CLI

Quartermaster1hr

Reviews and merges PRs

Lookout2hr

Scans production, handles deploys

Purser24hr

Audits deps, Firebase costs, bundles

Herald6hr

Writes tweets, posts to Discord for review

Every PR goes through the Quartermaster before merging. Herald writes @usdctofiat content and posts it to Discord for me to review before it goes live.

Ecosystem

Third-party skills and MCP servers I use alongside everything above.

  • VercelDeployment, performance, framework guidance
  • ImpeccableFrontend design, audit, polish
  • Firebase MCPDirect Firestore access from the agent
  • Chrome DevTools MCPHeadless browser for testing and QA
  • Context7Up-to-date library docs