A vendor-neutral spec for building real software with fleets of coding agents.
The Spec → · One-page gist → · Implementations → · vs Symphony / Spec Kit → · Sources → · The talk (slides) →
We thought AI would make us code faster. Instead, it changed what coding is.
Prompts don't remember. Single agents lose the plot. And "just add more agents" makes things worse: five chained steps at 95% reliability each succeed only 77% of the time; twenty agents drop you under 36%. Errors don't cancel out — they compound.
The fix isn't more agents. It's orchestration: one owner of context, isolated workers, no peer-to-peer chatter, deterministic gates. Orchestrated Coding is the name for that system, written down so you don't have to rediscover it.
This spec sits above single-agent specs. 12-Factor Agents tells you how to build one reliable agent; Orchestrated Coding tells you how to orchestrate many of them to ship a codebase.
Prompting → collapses at MEMORY (you become the context bus)
Single agent → collapses at CONTEXT LIMITS + NO INDEPENDENT CHECK
Naive multi-agent→ collapses at COMPOUNDING ERRORS + CONFLICTING EDITS
─────────────────────────────────────────────────────────────────────
Orchestration → the resolution: 3 Foundations + 7 Disciplines
3 Foundations — the substrate every fleet needs:
| Foundation | Fixes | |
|---|---|---|
| F1 | A single source of truth, treated as law (mechanically enforced) | reinvention / drift |
| F2 | Shared context as a protocol (read-before, write-after) | "prompts don't remember" |
| F3 | Deterministic gates as ground truth | "looks done but isn't" |
7 Disciplines — orchestration itself:
| Discipline | Fixes | |
|---|---|---|
| D1 | An orchestrator owns context & dispatch | dispersed decisions |
| D2 | Isolation per unit of work | conflicting edits |
| D3 | Handoffs as durable artifacts | context fragmentation |
| D4 | Dependency-ordered release (the gap loop) | wrong-order work |
| D5 | Independent review + deterministic merge gate | unreviewed output landing |
| D6 | Multiple harnesses, not multiple instances | single-model blind spots / stalls |
| D7 | Control loops = events wired to roles | manual babysitting |
Plus recovery as the property that matters most: how sticky failures are, not how often they happen.
Read the full spec, with conformance levels L0–L3 →
- L1 Orchestrated — foundations + orchestrator + isolation + handoffs.
- L2 Governed — L1 + dependency/gap loop + review/merge gate.
- L3 Resilient — L2 + multi-harness + control-loop automations + recovery + human-in-the-loop.
State your level in your README; link the evidence.
Itervox is an open-source orchestration daemon written in Go. It started as an implementation of the OpenAI Symphony spec and evolved into a full Orchestrated Coding runtime. It turns a Linear/GitHub backlog into a fleet of isolated worktree agents — with handoffs, a dependency/gap loop, multi-harness switching (Claude ↔ Codex), reviewer + merge-bot gates, and event-driven automations.
It conforms at L3. The spec's own audit caught two MUST-level gaps in it, and both were closed in v0.2.0 with named regression tests (see IMPLEMENTATIONS.md). Itervox is one way to build this — the spec is the point, the tool is the proof.
Building your own? Add it to the list →
No — it lives at a different altitude and composes with the specs you already know:
- OpenAI Symphony is the runtime spec for one orchestrator daemon.
- GitHub Spec Kit is the per-feature spec workflow.
- 12-Factor Agents is the single-agent spec.
Orchestrated Coding is the fleet layer above all three. It's designed against the MAST failure taxonomy, which found that ~79% of multi-agent failures are organizational — not model capability. Itervox implements two of the layers at once (a Symphony runtime and a clean OC L3), which is the proof that the layers stack. Full breakdown: COMPARISON.md.
The spec debuted as the conference talk From Prompting to Orchestrating: Coding Is Now a System. The full deck lives in slides/ as a runnable Slidev presentation:
cd slides && pnpm install && pnpm dev # opens http://localhost:3030Every number on a slide meets the same provenance bar as the spec — see slides/README.md.
Anyone running more than one coding agent and feeling the chaos: platform and AI-tooling teams, framework authors (Claude Code, LangGraph, OpenHands, Devin-likes), and engineers who've graduated from "prompt and pray" and want a system.
This is a v0.2 draft meant to be argued with. Propose a discipline, sharpen a normative line, or add your implementation with its conformance level. See CONTRIBUTING.md.
If this names something you've been struggling with — star it, and tell one person who's drowning in agents.