Web4Guru AI Operations
Docs · Specialists · CEO

The CEO Agent

The chief of staff. Your single point of contact. Takes your goal, designs the plan, spawns the specialists, reviews their output with the Evaluator, and reports the result back to you.

What the CEO does

The CEO agent is the 19th role on the team and the only one you talk to directly. It runs on the Claude Agent SDK as a long-running loop, holds a memory of your business, and decides — turn by turn — whether to answer you, spawn a specialist, run a cross-check with the Evaluator, or pause and ask you a question. The live source of truth for its behavior is apps/engine/src/ceo/loop.ts; the delegate tools it calls are in apps/engine/src/ceo/tools.ts.

When the CEO gets work

Every owner message enters through the CEO. There is no way to reach a specialist without going through it. When you sign in, the CEO is the loop the web app talks to on your behalf; when you close the tab, it keeps running until it reaches a decision point that requires you.

What it takes as input

A plain-English message from the owner. No structured brief, no YAML. Examples that work today:

  • “Build a landing page for my consulting service called Pam Strategy.”
  • “Research 20 chiropractors in Austin on Instagram — give me a ranked list.”
  • “Send the launch email to my list once I approve the draft.”
  • “Check my inbox and triage anything urgent.”

What it produces

Not artifacts directly — those come from specialists. The CEO produces decisions: which specialist, what brief, how to handle the verdict, how to present the result to you. The owner-visible output is a structured card (see structured cards) with the shipped deliverable, the workspace path, and a one-paragraph summary.

Tools it has access to

The CEO has the broadest tool belt on the team. From apps/engine/src/ceo/tools.ts:

  • delegate_to_coding_specialist, _content_, _research_, _browser_, _business_ops_, _design_, _product_, _chief_of_staff, _social_, _seo_, _ads_, _pr_, _sales_, _customer_success, _support_, _data_, _finance_, _legal_
  • spawn_parallel_specialists — fan out 2–5 independent tasks
  • delegate_sequence — chain specialists serially (Research → Content → Browser)
  • review_deliverable_with_evaluator — mandatory quality gate before surfacing a deliverable
  • emit_owner_card — the only way the CEO talks to you
  • Light built-ins: Read, Write, Glob, Grep, WebFetch, TodoWrite — used for triage, not execution
  • Memory tools: append_lesson, read_inter_agent_notes, request_replan

Workspace

The CEO's cwd is ~/.blackbox/ (or the configured $BB_DATA_DIR). Long-lived memory lives under ~/.blackbox/memory/: the business profile, lessons learned, inter-agent notes, evaluator verdicts. The CEO reads freely across every specialist's workspace but writes only to its own memory and to the shared task log.

Example exchange

Owner: “Build me a landing page for Pam Strategy — audience is SaaS founders, tone is sharp and opinionated, three sections: hero, problem, CTA. Have it live by tomorrow.”

CEO internal plan: (1) Use delegate_to_coding_specialist with task_id: "pam-strategy-landing" and the brief inlined. (2) While coding runs, use delegate_to_design_specialist in parallel for a moodboard. (3) Route the built page to the Evaluator with review_deliverable_with_evaluator. (4) If PASS, emit a report_ready card with the preview URL. (5) If the owner approves, proceed to deploy.

Owner-visible output

One card on your Dashboard titled “Landing page shipped — pam-strategy-landing”, with the preview URL attached, an Evaluator verdict in the body, and an Approve / Deploy button. No noise, no progress narration — you see the outcome.

Related specialists

The CEO works with every specialist. Most tightly: the Evaluator (mandatory review gate) and Business Ops (internal briefings, calendar prep, and the rest of the chief-of-staff work the CEO delegates rather than does itself). The full roster lives at /docs/specialists.

Frequently asked

Is the CEO a specialist?
Technically no — it is the orchestrator. But it belongs in this section because it is the role you talk to and the coordinator for everyone else.
Can it act without asking me?
For drafting, research, file writes: yes. For sending external communications, spending money, or deploying code: never without an approval card.
What stops a runaway loop?
The circuit breaker halts the session at five known safety conditions. No open-ended spend.

See also