C-Suite Manager Agent, Roles & Extensions

As autonomous tasks grow beyond single-file edits into multi-repository engineering campaigns and cross-department operations, manually prompting individual agents becomes a bottleneck.

Agent Harbor includes a persistent Manager Agent alongside declarative Roles (--role), Portable Extensions (--with), and Structured AI Workflows (ah spec, ah plan, ah implement, ah review).

Organizational Philosophy: One Human Owns Entire Departments

Agent Harbor is designed around two core operating principles:

1. One Human Per Department: Rather than coordinating dozens of humans across engineering, marketing, finance, and legal — where $O(N^2)$ meetings and handoffs bottleneck velocity — a solo developer or a small team assigns one human owner to an entire department. That human directs a fleet of concurrent agents through the Manager Agent, allowing the company to execute near the maximum speed the agents are capable of.
2. Async, Text-Driven Collaboration Mediated by Agents: Every specification, milestone transition, defect report, financial ledger, legal redline, and code review is recorded as a durable textual artifact (specs/*.md, milestones/*.org, issues/*.md, commits, and .ahr session recordings). Persistent oracle agents (@coordinator and @spec-index) and the Manager Agent continuously index these artifacts and synthesize high-signal reports for other humans on demand (ah session prompt @coordinator, ah spec ask), eliminating routine verbal status meetings.

For teams that want unattended closed-loop execution, Workspace Templates (ah workspace init --template <name>) provide both Department-Owner presets (solo-full-company, engineering-department, marketing-and-docs, finance-and-legal) and Zero-Human Autonomous presets (zero-human-company, zero-human-maintainer) where oracle question automation, multi-vendor consensus review, and hard budget circuit-breakers govern execution automatically.

Your Autonomous C-Suite

The Manager Agent operates as a coordinated C-suite of specialist executives that share organizational memory, decompose high-level objectives, dispatch specialist worker agents across your fleet, and verify deliverables:

Persistent Semantic Memory & Project Index

The Manager Agent continuously indexes your repositories, architectural specifications (specs/), active issues, pull requests, and historical .ahr session outcomes. When you ask a question or launch a workflow, it injects the relevant architectural context automatically so worker agents do not waste tokens rediscovering how your codebase is structured.

Structured AI Workflows (ah spec, ah plan, ah implement, ah review)

Instead of ad-hoc prompting, Agent Harbor provides structured workflow commands that coordinate specification authoring, planning, parallel implementation, and multi-agent review:

# 1. Research an architectural question or external ecosystem
ah research "Compare QUIC multipath implementations in Rust for fleet roaming"
 
# 2. Author or update a technical specification grounded in your codebase
ah spec "Add WebAuthn passkey support to the hub authentication service" --out specs/WebAuthn-Passkeys.md
 
# 3. Decompose a specification into an dependency-ordered execution plan
ah plan specs/WebAuthn-Passkeys.md --out plans/webauthn-plan.json
 
# 4. Execute the plan across one or more worker agents in isolated CoW branches
ah implement plans/webauthn-plan.json --parallel 3
 
# 5. Run a multi-agent code, security, and spec-compliance review on the resulting branch or PR
ah review --branch feat/webauthn-passkeys --against-spec specs/WebAuthn-Passkeys.md
 
# 6. Ask a quick codebase-aware question without starting a full editing session
ah ask "Where are mTLS client certificates validated in the hub ingest pipeline?"

Roles (--role) and Portable Extensions (--with)

Every agent session can be specialised at launch with a Role and one or more Extensions—regardless of whether the underlying agent is Claude Code, Codex, Gemini, Grok, or OpenCode.

Assigning a Role (--role)

A role configures the agent’s system instructions, tool permissions, verification checklists, and domain persona:

ah claude --role vp-engineering "Review the milestone plan for v2.4 and identify critical-path bottlenecks"
ah agy --role cmo "Draft the announcement guide and migration docs for Minimal-Impact Mode"
ah codex --role general-counsel "Audit all Cargo.lock dependencies added in the last 30 days for copyleft licences"

Attaching Extensions (--with)

Agent Harbor’s extension engine translates skills, prompt libraries, and MCP servers into the native configuration format of whichever agent you run (CLAUDE.md, .codex/, .gemini/, MCP JSON configs) inside an isolated temporary overlay:

# Attach extensions from GitHub, npm, PyPI, or a local directory
ah claude \
  --with gh:metacraft-labs/rust-architecture-skills \
  --with npm:@modelcontextprotocol/server-github \
  --with ./team-extensions/security-checklist \
  "Implement the webhook signature verifier"
 
# Manage installed extensions
ah extensions list
ah extensions install gh:metacraft-labs/rust-architecture-skills
ah extensions search postgres

Omnichannel Control: Chat Bridges, Desktop & Voice

You can converse with the Manager Agent and steer worker sessions from wherever you work:

Chat Bridges (ah manager join)

Connect the Manager Agent to your team’s messaging channels so you can assign tasks, receive completion summaries, and click Approve / Deny on Egress and Sudo Broker requests directly in chat:

ah manager start
ah manager join slack --channel "#agent-fleet"
ah manager join discord --guild "Engineering"
ah manager join telegram
ah manager join teams
ah manager join email --address [email protected]

Desktop App (ah desktop) & Voice Mode (ah voice)

# Open the native Agent Harbor Desktop App (macOS, Linux, Windows) with system tray HUD
ah desktop
 
# Start a hands-free bidirectional voice session with the Manager Agent
ah voice