The 18 specialists
The hands of your AI company. Eighteen specialists plus an Evaluator. The CEO delegates; they execute.
TL;DR
Each specialist is a narrow agent with its own system prompt, its own tool belt, and its own working directory. The CEO delegates work by opening a scoped sub-session; the specialist runs, emits results, and disappears. No specialist touches your inbox or your code without the CEO routing work to it on purpose.
Why specialization
A general model can do most things at 70%. A specialist model —
same base model, different system prompt, narrower tool belt —
does its one thing at 95%. The difference compounds: a Coding
Specialist that knows git, gh, Railway, and your
repo layout ships a landing page in ten minutes; a general
agent wanders for forty and still forgets to commit.
Specialization also makes safety simpler. If the Legal Specialist only ever drafts and always flags for human review, no amount of prompt injection can talk it into "just sending the NDA." The guardrails live in the role, not in the prompt.
The core six
These ship in every Black Box account today.
- Coding Specialist. Writes code, runs shell, operates git + gh, deploys to Railway and Vercel.
- Content Specialist. Copy, email sequences, long-form, voice-matched to your brand.
- Research Specialist. Web search, competitor analysis, structured summaries — cached per owner.
- Browser Specialist. Headless Playwright for real web tasks: forms, scraping, end-to-end QA.
- Business Ops Specialist. Email, calendar, invoicing, paperwork — the logistics layer.
- Evaluator. Independent reviewer for every specialist deliverable before it reaches you.
The expansion twelve
The rest of the roster. Some are live in gated beta; others ship on the expansion roadmap. Tier availability varies — see the pricing page for which plan unlocks what.
- Design Specialist. Layouts, brand assets, generated imagery, pitch decks.
- Product Specialist. Specs, user stories, roadmaps, release notes.
- Chief of Staff. Internal coordination, meeting prep, decision memos.
- Social Specialist. Platform-native posts, content calendars, engagement.
- SEO Specialist. Keyword strategy, on-page audits, schema, internal linking.
- Ads Specialist. Ad copy, audience targeting, A/B variants.
- PR Specialist. Press outreach, media lists, pitches.
- Sales Specialist. Prospect research, personalized outreach, pipeline updates.
- Customer Success. Onboarding flows, retention check-ins, QBRs.
- Support Specialist. Ticket triage, draft replies, knowledge-base drafts.
- Data Specialist. Analysis, dashboards, SQL, chart generation.
- Finance Specialist. Invoicing, P&L summaries, cash-flow tracking.
- Legal Specialist. Contract drafts, terms, privacy policies (always flagged for human review).
Anatomy of a specialist
Every specialist is three things stapled together.
- A system prompt. The role description, the do-not-touch list, the communication style. Written by Web4Guru, shipped with the engine.
- An allowed-tools list. A tight whitelist — e.g. the Coding Specialist has
Bash,Edit,Write, and Git + Railway CLI commands; the Legal Specialist has onlyReadand a document-generation tool. - A working directory. Per-task scoped filesystem —
~/.blackbox/coding/<taskId>/, for example. The specialist cannot reach outside its own cwd for files.
Skill packs add another layer on top — vertical expertise (restaurant ops, real-estate compliance, SaaS growth) wired into the system prompt and tool belt at delegation time. See skill packs.
Parallel vs sequential
Specialists run in parallel when the work is genuinely
independent (build the page AND write the launch email AND
research competitors) and in sequence when there is a dependency
chain (Research → Content → Browser to publish). The CEO picks
the shape by calling either delegate_to_* (one
specialist) or spawn_parallel_specialists (a fan-out
of branches that join back).
How it works in Black Box
Specialists live in apps/engine/src/specialists/<role>/.
Each folder exports a spec.ts (system prompt +
allowed tools), a delegate.ts (opens the
sub-session), and a workspace helper. The CEO
invokes them via MCP tools that wrap the Agent SDK's
query() call; each sub-session is independent at
the SDK level even though all specialists share the engine
process.
Every deliverable a specialist produces is reviewed by the Evaluator before the CEO accepts it. This is not a formality; the Evaluator has authority to send the work back with a critique and a re-do instruction.
Frequently asked
- Why eighteen and not one?
- Specialization beats generality at this scale.
- Are all eighteen shipped?
- Six core specialists today. The expansion twelve roll out in Pro, Scale, and Enterprise tiers on the expansion roadmap.
- Can I talk to a specialist directly?
- Not by design. You talk to the CEO.
- Can I bring my own specialist?
- Enterprise, yes. Public authoring is on the roadmap.
Related concepts
- The CEO agent — who delegates to specialists.
- Delegation — how the CEO chooses a specialist.
- Skill packs — vertical expertise you can attach to specialists.
- The Evaluator gate — who keeps specialists honest.
- Feature page: specialists.
- Glossary: specialist agent.
Next: how the CEO picks
Eighteen specialists and a goal. How does the CEO decide who gets the job?