Agentic Orchestration System

A config-driven framework that assembles specialized AI agents into deliberation councils and execution teams on top of Claude Code, Codex, Gemini, and Pi.

Get Started Explore Agents

$ bun add aos-harness
$ bunx aos init
$ bunx aos run strategic-council --adapter codex

# local install or global CLI. 15 agents. 9 profiles.

Agents

Profiles

Domains

Skills

Two Orchestration Patterns

Every decision is either a debate or a build. AOS handles both.

deliberation

Deliberation

A neutral Arbiter frames the question. Specialist agents debate from opposing perspectives. The Provocateur stress-tests consensus. Output: a structured memo with ranked recommendations and documented dissent.

Brief → Arbiter → Broadcast → Tension Pairs → Provocateur → Memo

Strategic Council, Security Review, Architecture Review

execution

Execution

A CTO Orchestrator drives the product lifecycle. Agents produce artifacts in sequence — requirements, architecture, planning, tasks, security review, and final assembly. Output: a complete execution package ready for engineering handoff.

Brief → CTO → Requirements → Architecture → Tasks → Package

CTO Execution Orchestration

Goes further with Dev Execution

Dev Execution goes beyond planning to actual code implementation. An Engineering Lead spawns domain-scoped worker agents that write code directly to your working directory via hierarchical delegation — from brief to committed-ready changes in a single session.

Learn about Dev Execution →

The Agents

15 cognitive agents, each with a distinct bias, temperament, and evidence standard.

Arbiter

Synthesize competing perspectives into actionable, ranked recommendations with documented dissent

orchestrator premium

CTO Orchestrator

Deliver a complete, buildable execution package with minimal ambiguity

orchestrator premium

Advocate (user-behavior-reality)

perspective ›

Architect (system-durability)

perspective ›

Catalyst (speed-and-monetization)

perspective ›

Navigator (positioning-and-timing)

perspective ›

Provocateur (truth-seeking)

perspective ›

Sentinel (sustainability-and-trust)

perspective ›

View all 15 agents →

Connect to the Agent Mesh

Don't stop at one runtime. Serve an AOS assembly as an Agent2Agent (A2A) agent, call other agents, and bind skills to MCP tools — with identity, egress, and ingress secured by default.

Serve over A2A

Expose your skills as a JSON-RPC A2A endpoint, skill-routed on a warm worker. aos serve --a2a.

Learn more →

Bind MCP tools

Map a skill to a Model Context Protocol tool for a native, deterministic call — no LLM round-trip.

Learn more →

Secured by default

Signed Agent Cards, JWKS key discovery, SSRF-screened egress, and rate-limited, owner-scoped ingress.

Learn more →

Supported CLIs

Bring the AI CLI you already use. Install the matching adapter, keep Pi pinned to explicit tier models, and let Codex, Claude Code, and Gemini use their vendor-default models unless you intentionally override them.

Pi CLI

Keep Pi installed, add the Pi adapter, and run the same AOS profiles through the Pi extension package.

$ npm i -g @aos-harness/pi-adapter

Claude Code

Use your existing Claude Code installation, add the adapter, then optionally install the project command pack. Readiness checks surface when ANTHROPIC_API_KEY is overriding Claude auth.

$ npm i -g @aos-harness/claude-code-adapter
$ bash plugins/aos-harness/claude-code/install.sh

Codex CLI

Pair Codex CLI with the Codex adapter, then optionally install the local Codex plugin bundle on top. By default, AOS defers model choice to Codex unless you pin a tier model.

$ npm i -g @aos-harness/codex-adapter
$ aos init --apply

Gemini CLI

Run the same deliberation and execution profiles through Gemini after installing the matching adapter package. Current account-level Gemini failures are surfaced directly instead of hidden behind stale flags.

$ npm i -g @aos-harness/gemini-adapter

Ready to orchestrate?

Install the framework. Write a brief. Let 15 agents debate your hardest question.

Get Started