walid@portfolio:~/lab/agency-agents-roster$
cd../lab
01ideaSep 2026

Two hundred specialists, install five

A roster of role-defined agents for your coding tool — and why the restraint matters more than the catalogue

Instead of one generic assistant, a library of specialists — each with a defined role, workflow, deliverables and success criteria — that install into your coding tool as agent files. The catalogue is the headline; the discipline is the point. Install three to five that match your actual week, give each a concrete deliverable, and chain their outputs rather than treating them as separate chatbots.

AgentsClaude CodeMITRosterOpen sourceThe repoDesktop app
i
Match the job to the specialist

The failure mode this is built against is asking one generic assistant to “build me a website”. Each agent here arrives with a role, a personality, a workflow, named deliverables and success criteria — so the instruction becomes “activate the frontend developer and build a responsive landing page”, and the output is judged against that role’s standard rather than general helpfulness. The specificity is doing the work, not the size of the catalogue.

Checked, not copied

The numbers, verified on 18 September 2026

Read from the repository itself rather than taken from the write-up.

ClaimVerified
230+ specialistsHolds — the repository states 230+ agents across every division in its own README.
Open sourceMIT licensed.
Popularity153,300 stars, and actively updated.
Tool supportFar wider than the guide suggests — sixteen targets, not the four commonly listed.
Setup

Two ways in

A desktop app for Mac, Windows and Linux browses the roster and installs without copying files by hand — the easier route if you are starting out. The script is the same thing with more control.

terminal11 lines
git clone https://github.com/msitarzewski/agency-agents.git
cd agency-agents

# everything, into one tool
./scripts/install.sh --tool claude-code

# or just the division you need
./scripts/install.sh --tool claude-code --division marketing

# see what is available before committing to anything
./scripts/install.sh --list teams
Wider than advertised

Sixteen install targets

The write-up names four. The installer actually supports all of these, which matters if your team is not standardised on one tool.

Targets
claude-code · codex · cursor · copilot · windsurf · aider
gemini-cli · qwen · kimi · hermes · opencode · openclaw
antigravity · osaurus · vibe · all
!
The gotcha the guide does not mention

On one of those targets the runtime registers only about 119 agents and silently drops the rest — an upstream bug, not a fault in the roster. Install everything there and you will have no idea which specialists are missing until one fails to respond. Installing a subset by division keeps you under the limit, and the installer warns you when a selection would exceed it. Worth knowing before you spend an evening debugging an agent that was never loaded.

The shape of it

Where the agents actually are

Counted from the repository. The distribution is lopsided in a useful way — engineering and the specialised catch-all are more than half the roster between them.

DivisionRoughly
Engineering62
Specialized59
Marketing36
Security12
GIS12
Design10
Testing9
Sales9
Project management, paid media7 each
Support, spatial computing, game development, academic6 each
The actual advice

Do not install everything

The catalogue is a menu, not a shopping list. Two hundred agents you never invoke is noise in the picker.

Start with three to five that match the work you are genuinely doing this month.Building a product: product manager, UX designer, frontend developer, backend architect.Growing an audience: content creator, SEO specialist, social strategist, copywriter.Building AI products: AI engineer, prompt engineer, data engineer, multi-agent architect.Add more only when a real task asks for one — not in anticipation of a task you might have.
Usage

Activate, do not ask

The difference between a vague request and a role-scoped one is the whole value of installing these at all.

prompts.txt10 lines
# instead of
Build me a website.

# do this
Activate Frontend Developer mode and build a responsive
landing page for my SaaS.

# or
Activate SEO specialist mode. Audit my website and give me
a prioritized 30-day SEO plan.
The power move

Chain them, do not silo them

Give each specialist a clearly defined job and pass its output to the next. This is the step that turns a folder of prompts into something that feels like a team.

handoff9 lines
Strategy Agent
   ↓  positioning + target
Product Manager
   ↓  requirements + scope
UI Designer
   ↓  layout + states
Frontend Developer
   ↓  implementation
Code Reviewer
i
Context plus a concrete deliverable

The single change that most improves output: name the role, give the situation, and demand a specific artefact. Not “help me with marketing”, but “you are the marketing specialist; I am launching an automation product for small businesses; analyse this landing page, identify the top five positioning problems, and return a revised positioning statement plus ten launch post ideas”. Specific inputs produce specific outputs, and a named deliverable is what makes the result checkable.

Worth knowing

Flags the write-up leaves out

The dry run in particular — previewing before writing a couple of hundred files into your agent folder is the sensible default.

terminal6 lines
--dry-run          # preview what would be installed, change nothing
--division <name>  # one division only
--agent <name>     # a single specialist
--agent-file <path>
--list             # what is available
--tool all         # every supported target at once
i
Provenance

The setup commands, activation examples and chaining pattern follow the source write-up. Everything else was read from the repository on 18 September 2026 and added here because the source omits it: the MIT licence, the 153,300-star figure, the full sixteen install targets against the four commonly quoted, the per-division agent counts, the dry-run and single-agent flags, and the runtime limit that silently drops agents on one target. The 230+ figure is the repository’s own and holds. Tracking parameters on the source’s links were stripped, and the author’s sign-off is not carried across.