Loomkin

Multi-agent platform built on Erlang/OTP (actually Elixir). Agents form teams, spawn specialists in milliseconds, share discoveries in real-time, review each other's work, debate approaches and vote on decisions, self-heal when things break, verify output before moving on.

Key differentiators from traditional AI assistants: - Teams-first: every session is a team of 1+ that auto-scales - Decision graph: persistent DAG of goals, decisions, outcomes (7 node types, typed edges, confidence scores) -- not just chat history - Context Mesh: overflow offloaded to Keeper processes with staleness tracking and failure memory, 228K+ tokens preserved vs 128K with zero loss - Agent spawn <500ms (GenServer.start_link), microsecond coordination via PubSub - 10 cheap agents for ~$0.25 vs ~$4.50 single-Opus - Self-healing: error classification, automatic diagnosis and repair via ephemeral agents - Verification loops: autonomous write-test-diagnose-fix cycles with upstream verifiers - Conversation agents: any agent can spawn brainstorms, design reviews, red team exercises - Speculative execution: agents work ahead on likely next steps

5 built-in roles: lead, researcher, coder, reviewer, tester. 58 built-in tools. 16 LLM providers, 665+ models. LiveView web UI with 39 components, zero handwritten JavaScript.

328 source files, ~77K LOC, 2,700+ tests.