Build to learn, not to sell. The code is open. If I recommend it, I'm using it.
Install · First mission · What you get · How it works · Commands · Docs
AGENT-11 deploys 11 specialist AI agents into your project and coordinates them through pre-built missions. Instead of prompting Claude Code task by task, you run /coord build and a coordinator plans the work, delegates it, checks the result against evidence, and keeps a project plan up to date across sessions.
It installs into an existing repo in one command and adds no runtime dependency: everything is markdown instructions, a settings file and a few shell scripts that Claude Code reads.
Why it's open: I build to learn, not to sell, and I give the code away. The value is the knowledge, the judgement and the trust, not the tool. If it is here, I am using it; if it broke, you will hear why. This is the squad behind my own builds (Trader-7 and the AI Search Mastery tools); the build story lives at jamiewatters.work.
Honest positioning. There are no benchmark numbers in this README because I do not have any I would stand behind. What I can tell you is what it does, how it is built, and where it breaks: see Where it struggles.
Requires Claude Code and a project directory with a git repo, a README, or a package file (package.json, requirements.txt, etc.).
cd /path/to/your/project
# Downloads, verifies its checksum, then installs
bash <(curl -fsSL https://raw.githubusercontent.com/TheWayWithin/agent-11/main/project/deployment/scripts/secure-install.sh)
# Restart Claude Code so it picks up the new agents
/exit
claudeVerify it landed:
ls .claude/agents/ # 11 specialists: analyst.md, architect.md, coordinator.md, ...
ls .claude/commands/ # 14 slash commands
ls missions/ # mission library, at the project rootNote the mission files are named mission-build.md, mission-fix.md and so on, but you invoke
them by the bare name: /coord build, /coord fix.
The same command upgrades an existing installation. Your own /CLAUDE.md at the repo root is never touched. AGENT-11 writes agents and commands into .claude/, and puts missions/, templates/ and field-manual/ at the project root; an existing .claude/CLAUDE.md is backed up before it is updated. Two-file architecture → · Upgrade notes →
Pick the one that matches where you are.
Existing codebase, want the squad to understand it:
/coord dev-alignment
Reads the repo, writes project-plan.md, architecture.md and a context file so later missions start informed.
New project, have an idea document:
/coord dev-setup ideation.md
Scaffolds the project and produces a plan you can start building against.
Just want to see it work:
echo "The login button does nothing on mobile. Expected: submits the form." > bug.md/coord fix bug.md
The shortest real mission. You will see the coordinator delegate, the developer change one thing,
and the tester verify with real command output before it is called done. /coord takes a mission
name and an optional input file; it will not infer a mission from free text, by design.
Missions write progress to project-plan.md as they go, so you can stop and resume. /coord continue picks up from the last passed phase gate.
| Agent | What it is for |
|---|---|
@coordinator |
Orchestrates missions, delegates, holds the phase gates. The one you talk to. |
@strategist |
Requirements, user stories, scope |
@architect |
System design, technology choices, API shape |
@developer |
Implementation |
@tester |
Test suites, edge cases, evidence-backed verdicts |
@designer |
UI/UX, design systems, accessibility |
@documenter |
READMEs, API docs, guides |
@operator |
Deployment, CI/CD, infrastructure |
@analyst |
Metrics, KPIs, data |
@marketer |
Positioning, launch, copy |
@support |
Bug triage, user feedback |
Call one directly with @developer add rate limiting to the login route, or let /coord route the work.
13 run with /coord <mission> [input-file]. /coord validates the name against a fixed routing
table and stops on anything it does not recognise; it will not guess. Durations are each mission
file's own estimate, not a measurement.
| Mission | Does | Typical |
|---|---|---|
dev-setup |
Greenfield project initialisation | 30-45 min |
dev-alignment |
Understand an existing codebase | 45-60 min |
build |
New feature, end to end | 4-8 hrs |
fix |
Bug resolution | 10-30 min scoped, up to 3 hrs |
refactor |
Code quality, no behaviour change | 15 min targeted, up to 4 hrs |
mvp |
Minimum viable product | 1-3 days |
deploy |
Production deployment | 1-2 hrs |
document |
Documentation | 2-4 hrs |
migrate |
System or platform migration | 4-8 hrs |
optimize |
Metric-driven ratchet optimisation | 2-6 hrs |
security |
Security audit and remediation | 4-6 hrs |
integrate |
Third-party integration | 3-6 hrs |
release |
Version release management | 2-4 hrs |
Five more ship but are not on the /coord routing table yet. They install, and you can run them
by opening the mission file and following it, or by handing it to the coordinator directly
(@coordinator run the mission in missions/mission-architecture.md). operation-recon also has its
own command, /recon. This gap is tracked as A11-ISS-17; it is stated here rather than papered
over, because /coord architecture currently returns an error rather than running anything.
| Mission | Does | How to run |
|---|---|---|
connect-mcp |
Discover and connect MCP servers | mission file |
architecture |
Architecture documentation | mission file |
product-description |
Product definition with risk register | mission file |
operation-genesis |
Full feature, concept to production | mission file |
operation-recon |
UI/UX reconnaissance | /recon |
Opinionated implementations Claude loads on demand: saas-auth, saas-billing, saas-payments, saas-email, saas-analytics, saas-onboarding, saas-multitenancy, plus code-review-loop, where a read-only critic scores a diff, a read-write fixer addresses only what was raised, and the critic re-audits until two clean rounds or a cap.
Browse and install the seven SaaS skills with /skills. code-review-loop installs with AGENT-11 and is not in that catalogue; read its SKILL.md and invoke it directly.
The coordinator is the loop. /coord starts a mission; /coord continue resumes it. It reads project-plan.md for state, delegates one task at a time to a specialist, and refuses to advance past a phase gate until that phase's criteria pass on evidence. It converges on two clean verification rounds rather than counting to a fixed number, and escalates to you when a phase burns its error budget.
State lives in files, not in the conversation. project-plan.md (what is done, what is next), progress.md (a changelog), and per-phase context files. That is what lets a mission survive a restart, and what lets you read what happened without scrolling a transcript.
Verification is default-fail. Every success criterion starts false and flips to true only when a command run block with real output is attached. "The code looks correct" is recorded as a guess, and a guess counts as a failure. This is the single behaviour that most changes the quality of what you get back.
The thing that judges the work is read-only to the thing doing it. Be precise about what that buys you, because the gap is where reward-hacking happens:
| What | Enforced? | By what |
|---|---|---|
.quality-gates.json anywhere in the tree |
Yes | Two of the four permissions.deny rules, plus the PreToolUse Bash hook |
gates/ and .gates/ at the repo root |
Yes | The other two rules, plus the same hook |
A gates/ directory nested deeper, e.g. packages/api/gates/ |
No | Edit(gates/**) is root-anchored. Add a rule for it if you have one |
| A test elsewhere that serves as acceptance criteria | No | Nothing. Agents are instructed not to edit it and generally comply; nothing refuses them |
| A benchmark or metric command that defines "better" | No | Nothing, unless you add an Edit(path) rule for it yourself |
The rules use the Edit(path) form, which Claude Code applies to Edit, Write, MultiEdit and
NotebookEdit alike. Write() and MultiEdit() rule forms are silently ignored, so do not write them.
The Bash hook is a speed bump, not a boundary: it catches redirection, tee, sed -i, cp, mv, rm, truncate, shred, unlink, dd of=, ln -s and in-place perl/ruby edits, but a write through an interpreter or a path held in a variable passes through, and no shell hook can catch those. The enforceable guarantee is the deny rules.
Quality gates guide → · Loop discipline guide →
Orientation is a stated rule, not a habit. Every specialist and mission carries the same instruction: Glob/Grep to locate before you Read, read only the lines you need, never read a whole file to find one symbol, do not re-read what you have already read. Finding what to change is the expensive step on a large repo, and this is the cheapest lever on what a session costs you.
| Command | Does |
|---|---|
/coord [mission] [file] |
Run a mission. /coord continue resumes; /coord alone lists missions |
/meeting @agent [topic] |
Conversational session with one specialist, no mission structure |
/plan |
Show or update project state |
/report [since] |
Progress report for a date range |
/skills |
Discover and install SaaS skills |
/foundations |
Process BOS-AI foundation documents into structured YAML |
/architect |
System design session |
/bootstrap |
Turn foundation documents into project-plan.md |
/design-review |
UI/UX review of pending branch changes |
/recon |
UI/UX reconnaissance sweep |
/pmd |
Post-mortem: analyse a failure, propose process fixes |
/dailyreport |
Daily progress report plus draft social posts |
/blog |
Draft a blog post from repo context |
/planarchive |
Archive completed work out of project-plan.md |
A good fit if you work solo or in a very small team, build with Claude Code already, want structure and an audit trail rather than one-off prompting, and are willing to read a plan file.
Not the right tool if you want a no-code builder, need a hosted service (this is files in your repo, nothing runs on my machines), work in a large team with an established process this would fight, or want an agent that ships unattended without you reading anything.
- Large codebases. Whole-repo analysis on a big project will exhaust context. Scope missions to a directory or a named surface.
- Long autonomous runs. Missions are designed to be watched, especially the first run in a new repo. Nothing auto-merges.
- Judgement calls. The coordinator escalates rather than guessing, which means it will stop and wait for you more often than a fully autonomous tool would. That is deliberate.
- The read-only gates do not cover everything, as the table above sets out. If a particular test or benchmark matters, add a deny rule for it.
- MCP servers are optional and unmanaged. If you connect them, credentials are yours to handle.
Re-run the install command from any project root. It merges settings rather than overwriting: your own keys win on conflict, hooks you wrote or edited are preserved, and only hook entries byte-identical to a version AGENT-11 shipped are updated.
Managing several repos? The bulk-ops toolkit under project/deployment/bulk/ applies an upgrade across a fleet in one pass.
Upgrade guide → · Release history → · Changelog →
Start here: Quick start · Installation · Troubleshooting
Working with it: Mission library · Agent roster · Field manual · MCP integration
Design and discipline: Quality gates · Loop discipline · CLAUDE.md architecture
Using it with BOS-AI: AGENT-11 is the technical execution arm of the BOS-AI ecosystem: BOS-AI plans the business, AGENT-11 builds the software. It works standalone. Handoff guide →
Issues and pull requests are welcome. The library is curated rather than open-by-default: if you want to add an agent or a mission, open an issue describing it and why it belongs before writing the PR, so the design gets discussed first.
If AGENT-11 has been useful, starring the repo helps other people find it, and a coffee is always welcome.
MIT. See LICENSE.
Built by Jamie Watters. Used daily on real products.