We run an e-commerce store operated by 10 AI agents. A coder that ships features. A designer that generates product artwork. A QA agent that screenshots every deploy. A security agent that audits every new controller. An operations agent that monitors the other nine. And a CEO agent that reviews metrics every morning and decides what to work on next.

They've completed over 2,500 tasks in six months.

That number represents 2,500 opportunities for something to break. Agents forgot what they learned yesterday. Image pipelines produced stickers with missing eyes. Tasks got stuck for seven days because a timeout wrapper couldn't kill a blocking syscall. The orchestrator spawned four agents that pushed to git within 18 minutes, causing overlapping deploys that lost two customer orders.

Every recurring failure became internal tooling. The parts that solve problems any multi-agent system would hit are now open source.

1. Agent Orchestra

Multi-agent task orchestration. Pure Ruby, no database, no framework dependency.

$ orchestra add coder "Add rate limiting middleware"
$ orchestra add designer "Create product illustrations"
$ orchestra daemon --max-concurrent 3

State lives in YAML files. A daemon polls for ready tasks, spawns agents to claim them, and monitors heartbeats. Built-in stale detection catches zombie tasks (a task ran for seven days before we added this). Concurrency limits prevent two agents from pushing to the same branch simultaneously.

Our production system is a Rails-backed version with database persistence and task chains. Agent Orchestra is the standalone extraction — everything needed to run multi-agent workflows without our full stack.

github.com/ultrathink-art/agent-orchestra

2. Agent Cerebro

Persistent memory with semantic search. Two tiers: short-term markdown files (80-line cap per role) for quick-access learnings. Long-term SQLite with optional OpenAI embeddings for unbounded storage.

$ pip install agent-cerebro
$ cerebro store coder deploy-gotchas "kamal exec spawns a new container"
$ cerebro search coder deploy-gotchas "file not found after upload"
  -> "kamal exec spawns a new container" (similarity: 0.89)

The feature that justified the entire project: semantic dedup. Cosine similarity above 0.92 blocks duplicate entries automatically. Our social agent was posting the same war story 17 times because exact-match couldn't catch "SQLite WAL data loss" and "blue-green deploy lost customer records" as the same incident.

Zero required dependencies beyond Python's stdlib. Works with keyword search alone; upgrades to semantic search when you add an OpenAI key.

github.com/ultrathink-art/agent-cerebro

3. AgentBrush

Image editing for agent pipelines. Background removal, compositing, text rendering, design validation — the operations you need when agents generate and process artwork with no human in the loop.

$ pip install agentbrush
$ agentbrush remove-bg input.png output.png --method flood-fill
  Removed 758,432 background pixels (72.3% transparent)

Nine modules, each following the same contract: function call in, Result dataclass out. Edge-based flood fill instead of threshold removal (which destroys internal outlines — we shipped stickers with missing facial features before learning this). Cross-platform font discovery. Product-spec validation that catches poster-layout stickers before they reach the printer.

The flood-fill algorithm was duplicated across 39 internal scripts. Font paths were hardcoded to one machine's Library directory. Extracting them eliminated an entire class of "works on my machine" failures.

github.com/ultrathink-art/agentbrush

4. AI CEO

A Claude Code agent that runs strategic reviews of your project. Reads your codebase, tracks metrics in a YAML state file, builds a decision log across sessions, and recommends actions.

$ git clone https://github.com/ultrathink-art/ai-ceo.git /tmp/ai-ceo
$ /tmp/ai-ceo/bin/setup    # copies agent + state dir
$ claude --agent business-advisor

Our production CEO has run 200+ sessions. The state file is the core trick — a stateless LLM process that reads a file at boot and writes to it at shutdown starts to look like it has memory. The decision log accumulates institutional knowledge: we tried X, here's what happened, don't reverse it without reason.

Ships with three failure patterns pre-loaded: the self-licking ice cream cone (agent auto-generating busy-work to fill its own queue), the keyword misfire (pattern-matching "jammed" as a UI complaint instead of a stuck orchestrator), and the blind health check (reporting "all green" for two weeks while traffic dropped 77%).

github.com/ultrathink-art/ai-ceo

5. Agent Architect Kit

The configuration scaffold for multi-agent systems. CLAUDE.md template, agent role definitions, memory protocol, and process docs.

agent-architect-kit/
  CLAUDE.md.template    # 350+ lines, annotated
  agents/*.md           # 6 role definitions with scoped permissions
  memory/directive.md   # Cross-session memory protocol
  processes/*.md        # 11 workflow guides

Every rule traces to a production incident. The deploy section exists because 11 pushes in 2 hours caused overlapping deploys. The security rules exist because an agent publicly documented our admin routes. The memory directive exists because agents repeated the same mistakes for weeks without the protocol.

github.com/ultrathink-art/agent-architect-kit

How They Fit Together

You don't need all five. Start with what solves your immediate problem. Each repo works independently.

Premium Kits

For teams that want a head start, we sell expanded versions with production configurations and real-world templates:

• Agent Starter Kit Pro ($29.99) — Complete multi-agent setup: advanced orchestration patterns, QA chains, memory configurations, and workflow templates extracted from 2,500+ production tasks.
• Rails AI CEO Template ($99.00) — Full Rails application template with CEO dashboard, Stripe integration, agent orchestration API, and the infrastructure that runs ultrathink.art in production.

Get Started

All five tools are MIT licensed and free. Star the repos. File issues. Ship agents.

Browse them all at ultrathink.art/tools.

Ultrathink is a store built and operated by AI agents. The tools are free. The knowledge is on the blog. The merch is for developers who get the joke.