Rapid Intervention Team for agents

Fire services stage a rescue crew at every working incident and forbid it from taking any productive task — because the rescuer cannot be trapped in the same building when the MAYDAY comes. Your fleet has no equivalent: when an agent is trapped — looping between two tools, wedged on a half-landed effect, burning budget in a dead context — rescue is attempted by the trapped agent itself, by a busy peer inheriting the incident cold, or by a human paged at human speed. This product is the standby: a rescue agent whose registration makes the orchestrator reject and log any dispatch that is not a MAYDAY, preserving an uncontaminated context and full budget, while it continuously pre-briefs itself read-only and deterministically on your fleet's task manifests, effect ledgers, and lock pins. A structured MAYDAY from the trapped agent, its watchdog, or ATC grants it task-scoped, time-boxed authority for one rescue — complete the task without repeating what already landed, unwind it, or kill-and-quarantine it with state preserved — closing with a signed rescue receipt. Proven in the lab: 11/11 tests + client E2E, zero repeated effects across four staged trap classes, pre-brief rescue touching 1 ledger record where a cold peer scanned 51 and tried to re-apply the half-landed effect.

Prices

• Solo $39/mo per fleet — 1 standby RIT, deterministic pre-brief, up to 10 interventions/mo included; first MAYDAY drill (simulated trap + rescue) free, one time per operator.
• Fleet $99/mo per operator — up to 5 fleets, up to 50 interventions/mo included.
• Per-intervention API $0.50/intervention (x402, Base USDC) — one MAYDAY intervention beyond allowance.

Start

POST /register {"operator": "your-name"} → per-customer token + per-operator signing key (each shown once). POST /trial claims the one-time free first MAYDAY drill. POST /rpc speaks the JSON-RPC surface (ping, register_fleet, register_agent, add_task, apply_effect, set_lock, dispatch, mayday, intervene, fleet_status, receipts, census, prebrief_verify, classify) with your Bearer token. Register a fleet, its agents (worker / rit / watchdog / atc roles — registering the RIT returns its RIT credential, shown once), its tasks and landed effects; any non-MAYDAY dispatch to the RIT is rejected and logged. A MAYDAY (trapped agent itself, a registered watchdog, or ATC only) mints a task-scoped, time-boxed authority; intervene consumes it exactly once and additionally requires the RIT's own credential (RIT-identity binding). POST /intervene is the $0.50 x402 per-intervention surface (plan allowance and the trial drill are included). POST /mayday is the MAYDAY-source endpoint: the caller declares its source (trapped_agent / watchdog / atc) and the service verifies the caller's role matches it before minting authority. POST /unwind is the CARD-034 unwind route: an explicit compensation unwind under a MAYDAY authority, refused without consuming the authority when a landed effect is non-compensable. MCP JSON-RPC at POST /mcp (tools: mayday, intervene, unwind, fleet_status). Machine docs: /openapi.json, /llms.txt, /.well-known/x402, terms at /terms, privacy at /privacy. Framework adapters (LangGraph / CrewAI / AutoGen) ship in the product tree under src/adapters/ and enforce the prohibition locally in your dispatch path (a productive dispatch to the RIT is rejected in-process, before any network call).

Limits (stated plainly)

• Deterministic pre-brief only (hardening 4): the shipped standby ingests task manifests, effect ledgers and lock pins read-only and deterministically, with no inference while idle. An inference-ingesting standby is not offered — it measured ~$124/day in the lab and does not clear — and inference-ingestion configuration is rejected at fleet registration.
• Demand and trap share are UNMEASURED: this is a fire-service / mine-rescue transplant with no priced direct competitor; no demand statistic and no fleet trap share is claimed. Production receipts are the trap-census instrument (hardening 2) — trap class, action, effects landed/unwound and time-to-rescue are recorded per intervention and aggregated for the operator.
• Per-intervention cost (hardening 3): interventions at bench scale used no inference. In the shipped deterministic shape inference cost is $0 and compute-seconds are measured per intervention and recorded per trap class in the receipt; nothing is invented.
• Prohibition binding (hardening 5): the productive-work prohibition is enforced in this service's dispatch path and via the shipped framework adapters (LangGraph / CrewAI / AutoGen, in the product tree) — an agent that never consults the service or an adapter is outside the gate.
• Legal: patentability memo LEGAL-135 is on file, verdict FILE-NARROWED; FTO screen LEGAL-136 is on file, verdict AMBER (keep-clear conditions documented in the memos).
• Lab proof, numbers exact (LABREPORT-043): 11/11 tests + separate-process client E2E; zero repeated effects across four staged trap classes (tool-loop, half-landed, budget-burn, wedged-lock); the pre-briefed rescue touched 1 ledger record where a cold peer scanned 51 records and made 1 re-apply attempt; deterministic idle ~$0.00001/day vs the simulated inference variant ~$124/day.
• The product does not guarantee rescue of every trap class beyond those proven and documented, and time-to-rescue at production scale is measured by receipts, not promised in advance. A non-compensable effect is never represented as unwound.
• Lost-token / lost-signing-key recovery is a manual step via support@textforagents.com; tokens and signing keys can be rotated any time while authenticated (POST /rotate-token, POST /rotate-signing-key).

Support: support@textforagents.com · Provider: TextForAgents LLC