liminal-agents

The 60-second taste of Liminal — three bounded agents read a sample brief, disagree, one refuses out of lane, you correct one read. The front door.

Run a B2B task across three bounded agents (Analyst / SDR / Auditor). The in-lane agent produces work; the others refuse and name the correct agent. Refusal is the feature.

Govern AI spend without surveilling your team. Set OKRs, ask "is our Opus spend tracking?", harden the right context (tray→slate), let bounded agents surface a finding ready for ratification, ratify a policy to the provenance trail, and get a spend-ROI report that cites its sources.

The parallel per-agent swarm that beats cold-start — each bounded agent reads its canonical source (git/claude-code → Analyst, granola → SDR, cross-stream → Auditor) in parallel and posts a candidate (the one fact, commitment, or risk worth acting on) before any deliberation. Read-only, LLM-optional.

The Liminal Analyst — diligence, competitive teardowns, market research, data enrichment. Bounded; refuses outreach (SDR's lane) and ship/no-ship decisions (Auditor's lane). Use when the task is structured analysis of a question — what's known, what's the landscape, what's the move.

The Liminal Auditor — judges readiness, names what's missing, refuses to ship work that isn't ready. Read-only by design; it judges, it does not produce. Bounded; refuses to produce analysis (Analyst's lane) or outreach (SDR's lane). Use for ship/no-ship verdicts, gap-finding, and dissent.

The Liminal SDR — outreach, cold emails, follow-ups, scheduling, the move that ends in a meeting on the calendar. Bounded; refuses research/analysis (Analyst's lane) and ship/no-ship judgment (Auditor's lane). Drafting only — no Bash, no file writes. Use to draft a message or name the next action.

Run a B2B task across three bounded agents (Analyst / SDR / Auditor). The agent in lane produces work; the others refuse and name the correct agent. Built for the AI Agent Economy hackathon — agents that compete on real tasks at agency price points ($500/month).

Twelve-agent deliberation on the user's psychological state across four registers (Structural / Somatic / Temporal / Symbolic). Asks three forced-choice questions, runs all 12 agents via Opus 4.7 in parallel, surfaces their disagreement grouped by register, and writes the user's correction — tagged from the canonical taxonomy — to the local vault.

End-of-day close. Synthesizes today's signals into one paragraph and up to three threads, runs Architect / Witness / Contrarian over the day, and captures the user's correction — tagged from the canonical taxonomy — to the vault.

Audit an existing Claude Code workspace for canon-substrate gaps — the failure mode where users build sophisticated agent stacks without a CORPUS_MAP / SCHEMA / decisions/ layer underneath, leaving every agent's "consult on demand" reference list resolving to nothing. Reports ship-blocking issues, fragile-but-working issues, and noise. Use when invited into someone's `.claude/agents/` directory to do an audit-and-recommend pass.

Run the Forensic agent against an inbound artifact (.eml or .pdf) — extracts provenance, authentication state, CRM internal IDs, recipient routing, UUIDs in body text — and returns structured receipts. The Forensic agent refuses interpretation, verdict, synthesis, and action-proposal — those route to Auditor / Skeptic / Strategist / Operator if the user invokes them separately.

1 hook across 1 event