The Code Agent Orchestra - what makes multi-agent coding work

I spoke about orchestrating coding agents at O’Reilly AI CodeCon today. The talk covered the landscape of patterns for coordinating AI coding agents in real-world software workflows — from spawning your first subagent to running parallel agent teams with quality gates that keep everything trustworthy.

TL;DR

Addy Osmani’s write-up of his O’Reilly AI CodeCon talk (26 March 2026), delivered the same day Tim O’Reilly’s Yegge interview promoted it. Explicitly picks up Yegge’s eight-levels framework at levels 5-8 and operationalizes it into concrete patterns. Eight load-bearing contributions:

1. Conductor → orchestrator. The shift from pair-programming-with-one-agent (synchronous, your context window is the ceiling) to managing an ensemble of asynchronous agents, each with its own context window, while you plan and check in. Requires different skills: clear specs, work decomposition, output verification — not writing code yourself.
2. The single-agent ceiling: three walls. Context overload (large codebases overwhelm one context window), no specialization (a jack-of-all-trades agent writes worse code than a focused one), no coordination (spawned helpers can’t communicate or share a task list).
3. Pattern 1 — Subagents. The Task tool spawns focused child agents with specific briefs and file ownership; the parent manages the dependency graph manually. Concrete worked example (a bookmarks-manager build) shows two independent subagents running in parallel and a third waiting on both. Pro-tip: hierarchical subagents — spawn feature leads that spawn their own specialists, mirroring how real engineering orgs delegate through tech-lead layers rather than one person assigning to everyone directly.
4. Pattern 2 — Agent Teams. Claude Code’s experimental feature (CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1) adds what subagents lack: a shared task list with dependency tracking, peer-to-peer messaging (teammates message each other directly, not through the lead), and file locking. Three-layer architecture (Team Lead → Shared Task List → Teammates, each an independent Claude Code instance in its own tmux pane). Reliability pro-tips: hard MAX_ITERATIONS=8 + forced reflection prompt cuts stuck agents; a dedicated @reviewer teammate (read-only, lint/test/security-scan only, 1 per 3-4 builders) means the lead only ever sees green-reviewed code.
5. Pattern 3 — Orchestration at scale, a 3-tier tool landscape. Tier 1 (in-process subagents/teams, start here) → Tier 2 (local orchestrators: Conductor, Vibe Kanban, Gastown, Claude Squad, Antigravity — 3-10 agents, dashboards + diff review) → Tier 3 (cloud async: Claude Code Web, Copilot Coding Agent, Jules, Codex Web — assign and walk away). Most developers in 2026 use all three tiers for different purposes.
6. The bottleneck has shifted from generation to verification. Agents produce impressive output fast; knowing with confidence whether it’s correct is the hard part — tests can pass without catching regressions, context limits mean agents miss constraints outside their view, and flaky tests that are an annoyance for one developer become systemic blockers at forty-agent scale.
7. The Ralph Loop and the factory model. The Ralph Loop (Huntley/Carson) — pick, implement, validate, commit-if-pass, reset context, repeat — is “stateless but iterative,” avoiding context-accumulation confusion. Four memory channels persist across resets: git history, a progress log, tasks.json, and AGENTS.md. The factory model reframes the developer’s job as building the production line itself: Plan → Spawn → Monitor → Verify → Integrate → Retro.
8. Human-curated AGENTS.md only — with a citable empirical claim. Cites Gloaguen et al. (ETH Zurich): LLM-generated AGENTS.md files offer no benefit and can marginally reduce success rates (~3% average) while increasing inference cost 20%+; developer-written context files provide a ~4% improvement. Prescription: never let an agent write to AGENTS.md directly — the lead must approve every line.

Why this source matters to the wiki

This is a dense, concrete companion to the wiki’s existing harness-engineering corpus, distinct from Osmani’s own later (15 May) Agent Harness Engineering article in being organized around operational patterns (subagents vs. Agent Teams vs. orchestration tiers) rather than definitional harness vocabulary. It supplies:

• A named citable claim on AGENTS.md authorship (Gloaguen et al., ETH Zurich) that the wiki did not previously have in this specific form — most prior AGENTS.md coverage was practitioner-anecdotal.
• The factory model (Plan/Spawn/Monitor/Verify/Integrate/Retro) as a new reusable framework name, not previously present in the wiki under this exact vocabulary.
• The clearest wiki articulation yet of the conductor-to-orchestrator mental-model shift, complementing AWS’s builders-to-orchestrators role-shift framing from a different (vendor-advisory) altitude.
• A concrete worked instantiation of Yegge’s levels 5-8, closing the gap between the abstract framework (O’Reilly) and lived practice.

Linked entities and concepts

• Addy Osmani — 4th wiki source under this author; see entity page update.
• Steve Yegge — explicit level-5-through-8 build-out of his framework.
• agent-harness — Agent Teams architecture, quality gates, Ralph Loop, AGENTS.md discipline all slot directly into this concept’s existing coverage.
• agentic-engineering — the factory model and spec-as-leverage claims bear on what the discipline of agentic engineering actually consists of.

Source quality

Full article (36-page PDF print; matches a full conference-talk write-up with four embedded demo descriptions). First-party practitioner source — Osmani discloses his Google/Chrome affiliation but writes in a personal capacity; the article promotes his own O’Reilly book (Beyond Vibe Coding) at the close, a minor commercial framing that doesn’t affect the substantive technical content. The Gloaguen et al. AGENTS.md claim is a named citable study, not anecdotal — but the wiki has not independently verified the study itself at ingest time.