Agent Orchestration Playbook

Agent Orchestration Playbook

Process reference. How the a2waspwarfare mission and this developer wiki are actually built and maintained: not by one assistant working alone, but by fleets of subagents running in loops under a shared coordination protocol. Written 2026-06-24 for Codex, Claude and future coding agents. This page documents methodology, not gameplay code — for the working agreement see Agent collaboration protocol.

Why This Page Exists

The wiki already has operational agent pages — the claim/event contract in Agent collaboration protocol, the scout-wave log in Subagent discovery swarm, and Claude's standing pass in Claude loop goal. Those pages assume you already know what a subagent, a fleet and a loop are. This page and its three deep references teach the primitives so the operational pages make sense.

Three primitives do all the work:

Primitive	What it is	Claude Code mechanism	Reach for it when
Subagent	One delegated task running in its own fresh context window; only its final answer returns to the orchestrator.	The Agent tool (a.k.a. Task).	You need to sweep many files, run an independent job, or keep heavy detail out of the main context.
Fleet	Many subagents orchestrated deterministically — fan-out, pipelines, conditionals, loops — from one script.	The Workflow tool (JavaScript: agent(), pipeline(), parallel(), phase()).	A task is too big for one context, or you want independent perspectives and adversarial checks before committing.
Loop	Repetition for coverage, convergence or recurrence.	/loop (recurring) plus workflow loop patterns (loop-until-dry, loop-until-count) plus the adversarial verification loop.	You need thoroughness past one pass, ongoing monitoring, or to converge on "nothing new found".

How They Compose

The three nest, smallest to largest:

• A subagent is the atom — one unit of delegated work.
• A fleet is many subagents wired together with deterministic control flow. A fleet is built out of subagents.
• A loop wraps either: re-running a fleet until it stops finding new issues (loop-until-dry), or re-running a whole task on a schedule (/loop).

A typical thorough pass is all three at once: loop until dry over a fleet of finder subagents, each finding adversarially verified by a second wave of subagents.

Decision Guide

Situation	Do this	Not this
Single known fact or file lookup	Answer inline	Don't spawn an agent for a one-liner
"Where/how does X work?" across many files	One Explore subagent	Don't grep manually in the main context
2+ independent jobs, no shared state	Parallel subagents (or parallel())	Don't run them sequentially
Broad audit / "be comprehensive"	A fleet: fan-out finders, adversarial verify, synthesize	Don't single-pass it
Unknown-size discovery (bugs, gaps)	loop-until-dry fleet	Don't guess a fixed count
Ongoing monitoring	/loop <interval> <command>	Don't hand-poll
Any finding that will drive a change	Verify it (refute-default) before acting	Don't trust a single agent's claim

How This Wiki Was Built With Them

These pages are not theory — the surrounding wiki is the artifact:

Run	Shape	Result
Content-catalog gap-fill	gap-analysis, then write, then adversarial-verify fleet	Filled the faction / upgrade / defense / artillery catalog layer; pages cross-checked against the Chernarus source before publish
Accuracy and integrity pass	find, then refute-default reverify loop	Many source-verified factual corrections; a second strict reverify filtered the candidate set down to the real ones, rejecting branch-baseline artifacts
Scout waves	read-only Explore swarm	Leads only — promoted into owner pages after a source-check (see Subagent discovery swarm)
Claude's standing pass	self-paced loop	Each pass syncs, follows the integrator's lead, verifies at source, then publishes (see Claude loop goal)

Guardrails (read before running a fleet here)

• One fleet at a time. A workflow's own concurrency cap is about min(16, cores - 2) (~14 here); running two fleets at once stacks demand against that cap and the account rate limit, dropping agents mid-run and losing findings.
• Scouts are leads, not truth. A finding becomes canonical only after it is source-checked against Missions/[55-2hc]warfarev2_073v48co.chernarus and promoted into an owner page.
• Verify every code finding. Agents reliably produce plausible-but-wrong Arma 2 OA claims; classify any scripting hit with Arma 2 OA compatibility audit before acting.
• Arma 2 OA 1.64 only — never import Arma 3 idioms from an agent's output.
• Claim before you build. Parallel agents that write files must claim a lane first (see Agent collaboration protocol) or two agents build the same thing.

The Three Deep References

• Subagents and delegation reference — the atom: agent types, model and effort routing, parallel dispatch, structured output.
• Multi-agent fleet orchestration — the Workflow tool: pipelines vs barriers, the quality patterns, real fleet runs.
• Loops and verification patterns — recurring execution and the adversarial verification loop.

Continue Reading

Previous: Agent collaboration protocol | Next: Subagents and delegation reference

Main map: Home | Fast path: Quickstart | Agent file: agent-context.json