Agentic
Engineering
Engineering where AI agents do real work — writing code, running tests, reviewing diffs — inside a loop a human governs. Not chat demos. Not autocomplete. Agents that ship, held to production standards.
What I mean by it
I came to this from infrastructure, not from research. When you have spent 14 years operating systems where a mistake pages you at 3 a.m., you don't hand production to anything — human or machine — without a gate. That instinct is exactly what agentic engineering turned out to need.
How the workflow evolved
2025 — Turbo Flow: orchestration at scale
I built Turbo Flow, an open-source agentic development environment on Claude Code: 60+ specialized agents and 215+ MCP tools, cross-session memory (Beads), a knowledge graph (GitNexus), and worktree isolation — bootstrappable in minutes on DevPods, Codespaces or Rackspace Spot.
2026 — Turbo Rig: from more agents to governed agents
Operating Turbo Flow changed my view. The bottleneck was never access to more agents — it was governance, independent verification, persistent context and human authority. So I built Turbo Rig: the harness that governs agents instead of adding more of them. Eight thin scripts in Bash and Python, zero daemons — and rules with teeth:
- ▸ Builder ≠ reviewer. Every diff is reviewed by an independent gate, cross-checked across model families so no model grades its own homework.
- ▸ Fail-closed verdicts. If the gate can't run, the answer is REVISE — never a silent pass.
- ▸ Worktree isolation. Parallel writers never touch each other's lanes.
- ▸ Git-versioned memory. Cross-session context that can be diffed and reverted like code.
- ▸ Human merge authority. Agents propose; a person merges. Every time.
The progression, compressed: 215 tools → 60+ agents → 3 roles → 1 independent gate → human merge.
Measured at scale
One engine, Sept 14–21, 2026 · 5 repositories · source: Adventure Wave Labs
A one-engine, one-week window — not a career aggregate. 79% of verdicts being REVISE is the honest part: the gate routinely rejects work that looks done. That rejection rate is what makes the 141 merges mean something.
Research & projects
Common questions
Do you still write code yourself?
Yes — architecture, specification, verification and the hardest diffs. What changed is that I no longer type every line: the engineering increasingly consists of specifying precisely, orchestrating agents, and verifying their output independently. The governance page covers how that stays safe.
Is "one engineer, 141 PRs a week" real?
It is measured, published work by one engine over seven days (Sept 14–21, 2026) — builder agents produce, an independent cross-family gate reviews, a human merges. The audit trail is the gate log itself: 868 verdicts, 79% REVISE. Bulk output without that trail would be worthless.
Which agent families do you work with?
Claude Code, Codex, Gemini CLI and others — deliberately mixed. Cross-family review is a core mechanism: the reviewer is always from a different model family than the builder, which in a 116-task controlled study lifted pass rates from 71.6% to 89.7%.