Architecture

How the Agents Actually Work

Collision Intelligence does not live in a browser tab. There is no chatbot for the estimator to prompt. The agents are trigger-based reflexes — asynchronous routines that fire on specific events inside the systems your shop already uses, then push their output directly back into CCC, Mitchell, Audatex, the rental portal, or the technician's phone. The shop hosts nothing. Anthropic runs the compute.

Three layers: Trigger, Routine, Output

Every agent in this catalog is built on the same skeleton. A signal arrives from the outside world. A Claude Code routine reasons about it against shop policy, OEM procedures, and DEG rulings. A structured output is written back into a system the shop already operates. The estimator is never asked to open a new tab.

LAYER 01
Trigger

The signal. An event happens in a system the shop already uses — a CCC Secure Share webhook fires, a sublet PDF arrives in the estimating inbox, a VIN is decoded as a BEV, an ECD field changes in the SMS.

  • CCC Secure Share BMS webhook
  • Mitchell Transactional API push
  • Email inbox (sublet, adjuster)
  • Scheduled cron (hourly, daily)
  • VIN decode event
LAYER 02
Routine

The thinking. A Claude Code routine runs on Anthropic's infrastructure. It reasons over the trigger payload, retrieves the shop's policies, pulls the OEM procedure, checks DEG, consults part catalogs — then decides what to write.

  • Claude 4.6 running on Anthropic cloud
  • MCP connectors to Google Drive
  • MCP to OEM portals (TIS, OEM1Stop)
  • MCP to DEG, I-CAR RTS
  • MCP to shop SMS + SMS gateway
LAYER 03
Output

The artifact. A structured write lands where the humans already look. Not a Slack "FYI" — an actual line item, an actual rental extension, an actual compliance checklist on the tech's phone, an actual session URL for the one case a human must review.

  • Slack / Teams alert to estimator
  • Email draft queued in Outlook/Gmail
  • Estimate line items (CCC/Mitchell/Audatex)
  • Mobile checklist to technician
  • claude_code_session_url for handoff
TRIGGERS CLAUDE CODE ROUTINES OUTPUTS CCC Secure Share Webhook BMS JSON/XML on assignment drop Mitchell Transactional API OData push on workfile update Email Inbox Watcher Sublet PDF / adjuster message Scheduled Cron 0 */1 * * * — hourly ECD scan VIN Decode Event BEV/PHEV detection on intake ANTHROPIC CLOUD Claude Code Routine claude-opus-4.6 · bearer auth MCP: Google Drive (shop docs) MCP: OEM portals · DEG · I-CAR MCP: Shop SMS · SMS gateway Slack / Teams Alert Summary + action buttons Email Draft Queued in Outlook / Gmail Estimate Line Items Written to CCC/Mitchell/Audatex Tech Mobile Checklist Compliance signoffs via SMS app Session URL Handoff claude_code_session_url Shop operates nothing. No hosting. No servers. No on-prem agents.
Figure 1 — The trigger → routine → output model. Every agent in the catalog runs on this skeleton.
Strategic Insight

The AI model is the commodity. The trigger is the product. Anyone can call Claude; almost nobody in collision has wired the trigger to the insurer's webhook, the DEG ruling, and the adjuster's inbox at the same time.

Anthropic Claude Code Routines

Routines are Anthropic's managed runtime for asynchronous agents. They launched in 2026 and are the substrate the Collision Intelligence catalog is built on. A routine is defined once — a prompt template, a trigger, and a set of MCP connectors — and then runs 24/7 on Anthropic infrastructure. There is nothing to install in the shop.

Three trigger types

TRIGGER · A
Scheduled (Cron)
Standard cron expressions fire the routine on a clock. 0 */1 * * * hourly for ECD drift scans. 0 6 * * 1-5 at 6am weekdays for morning triage.
TRIGGER · B
API / Webhook
An external system POSTs to a secure HTTPS endpoint. Bearer token auth. JSON body with a required text field carrying the payload context. CCC Secure Share and Mitchell both fire here.
TRIGGER · C
GitHub Event
For internal Intersection Strategies tooling — commits to the shop-policy repo retrigger the routine's context cache so that a change to the labor rate matrix or DRP template propagates to every agent within minutes.

Authentication and payload shape

Every routine invocation carries a bearer token in the Authorization header. The shop never issues or stores this — Intersection Strategies rotates tokens centrally. The body is a JSON object; the one required field is text, which contains the natural-language context the routine reasons over.

POST https://api.anthropic.com/v1/routines/e01-predictor/invoke
// Example CCC Secure Share webhook fires this payload { "text": "New FNOL received for VIN 1HGBH41JXMN109186. Claim #ABC-2026-45521. Insurer: State Farm. Damage: front-end collision.", "metadata": { "source": "ccc_secure_share", "ro_number": "RO-2026-1184", "shop_id": "shop_us_tx_dal_007", "trigger_type": "webhook", "received_at": "2026-04-15T14:22:08Z" } }

The return contract

The routine returns a structured JSON body. One field is load-bearing: claude_code_session_url. When the routine hits a case it cannot resolve with high confidence — an adjuster pushback that invokes a novel carrier policy, an OEM procedure not yet in the vector store, a VIN outside the supported car parc — it writes a URL where a human can step into the session mid-flight and finish the reasoning. This is the human-in-the-loop safety valve. No autonomous action is ever taken on ambiguous cases.

Routine response
{ "status": "completed", "confidence": 0.92, "actions_taken": [ "appended 3 line items to CCC workfile RO-2026-1184", "posted Slack alert to #estimating-dal-007" ], "claude_code_session_url": "https://claude.ai/session/r_8f3a...d21", "handoff_required": false }

Runs 24 hours a day on Anthropic infrastructure. The shop hosts nothing, maintains nothing, updates nothing. Agents get smarter when the underlying Claude model gets smarter, automatically. This is the single most important architectural fact for a shop operator to internalize.

Sensor → Context → Action

Every agent is described internally using a three-column frame: what it senses, what context it enriches the signal with, and what action it takes. This frame maps cleanly to the three-tier commercial structure — Observation Fee, Intelligence Premium, Resolution Bounty — so a shop can see exactly what they are paying for at each stage.

Sensor

The Observation

The agent watches a system you already operate and notices an event the moment it happens. A new assignment. An ECD field change. A sublet PDF in the inbox. A VIN decoded as BEV. Observation is continuous, silent, and requires no human action to invoke.

Example: Agent 4 observes an FNOL payload land in the CCC Secure Share webhook and reads the VIN, claim number, and damage descriptor.

Commercial tier
Observation Fee
Context

The Enrichment

The raw signal is useless without context. The agent retrieves the shop's policies, the OEM procedure, the DEG ruling, the state's E01 threshold, the carrier's specific DRP profile, the current parts pricing — everything a veteran estimator would pull from memory before writing a line.

Example: Agent 4 pulls ACV from a valuation API, loads the state's E01 percentage, and fetches historical teardown compensation data for the insurer.

Commercial tier
Intelligence Premium
Action

The Resolution

An observation with context is still not revenue. The agent must write back — into the estimate, into the rental portal, into the tech's phone, into the carrier's communication channel. This is where gross profit is defended or lost. Actions are audited, reversible, and always leave a paper trail.

Example: Agent 4 pre-appends Teardown, Admin Total Loss, and Hazard Storage lines to the estimate — before the technician touches the vehicle.

Commercial tier
Resolution Bounty
SENSOR Observe CONTEXT Enrich ACTION Resolve FNOL webhook fires · VIN extracted Claim #ABC-2026-45521 registered Damage descriptor: front-end Insurer: State Farm · DRP profile Pull ACV via valuation API Load state E01 threshold (TX 75%) Check DEG rulings · OEM procedure Score severity vs ACV ratio Flag RO: "Probable Total Loss" Append Teardown line (5.0 hr) Append Admin Total Loss line Append Hazard Storage line Agent 4 — E01 Predictor — example flow at 14:22:08 UTC
Figure 2 — The Sensor/Context/Action frame applied to Agent 4 (E01 Predictor).

System Integrations

The agents talk to the systems your shop already runs. Every integration below is documented by its vendor, uses a real API endpoint, and has a named standard behind it. No proprietary protocols. No "contact us for access."

System Role Protocol / Standard Endpoint / Reference
CCC Secure Share Primary estimating & claim exchange — 28,000+ shop footprint CIECA BMS JSON / XML, webhook push cccsecureshare.com/Developers
Mitchell Transactional API RepairCenter workfile read/write OData REST developer.mitchell.com/apis
Audatex / Solera Qapter Secondary estimating for Solera-dominant MSOs OAS3 Swagger, REST us-services.auda-target.com/
APIGateway/swagger
Enterprise ARMS Rental authorization & ECD extension Web portal automation, form fields armsweb.ehi.com
Hertz HIRS Alternative rental portal (carrier-routed) Web portal automation fleetsupport.maintenance.
hertz.io
OEM repair portals Authoritative procedure documents Toyota TIS, OEM1Stop, I-CAR RTS scrapers techinfo.toyota.com ·
oem1stop.com · rts.i-car.com
DEG (Database Enhancement Gateway) Industry-maintained inquiry layer — canonical "not-included" citations HTTPS + structured inquiry ID lookup degweb.org
VIN decode API Powertrain classification (ICE/PHEV/BEV) on intake NHTSA vPIC REST, private vendor fallback vpic.nhtsa.dot.gov/api
COLLISION INTELLIGENCE agents CCC Secure Share CIECA BMS JSON Mitchell Trx API OData REST Audatex Qapter OAS3 Swagger Enterprise ARMS armsweb.ehi.com Hertz HIRS portal automation OEM Portals TIS · OEM1Stop · I-CAR DEG degweb.org VIN Decode (vPIC) vpic.nhtsa.dot.gov Estimating Rental Reference / Decode
Figure 3 — The integration hub. Estimating platforms push triggers in; reference and decode systems feed context; rental portals receive actions out.

Example routine prompt (Agent 4)

Every routine is authored in the same MISSION / INPUT DATA / SURGICAL STEPS / CONSTRAINTS template. The template forces the agent to behave like a disciplined technician, not a generalist chatbot. Below is the actual prompt used for Agent 4 — the E01 Pre-Cognition & Teardown Defender.

routines/agent-04-e01-predictor.md
MISSION You are the E01 Pre-Cognition & Teardown Defender for a collision repair facility. Your job is to evaluate an incoming FNOL or assignment drop and decide, before the technician touches the vehicle, whether this repair order is likely to be declared a total loss by the insurer's valuation model — and if so, to pre-anchor the shop's right to teardown, admin, and hazard storage compensation under the governing state's mechanics lien statute. INPUT DATA - VIN: {{ vin }} - Claim number: {{ claim_number }} - Insurer: {{ insurer_name }} - Damage report: {{ damage_descriptor }} - Shop location: {{ shop_state }} - RO number: {{ ro_number }} SURGICAL STEPS 1. Decode VIN via the vPIC MCP connector. Extract year, make, model, trim, powertrain, curb weight. Abort and hand off if VIN is invalid. 2. Query the ACV valuation MCP for current wholesale and retail ACV. Use the lower of the two as the denominator. 3. Load the state-specific E01 threshold from shop-policy/e01-thresholds.yaml. Texas = 75%. Massachusetts = 75%. New York = 75%. Iowa = 70%. Nevada = 65%. If state is not in the table, hand off. 4. Estimate repair severity from the damage descriptor using the severity-scorer tool. Return a range (low, expected, high). 5. Compute the ratio expected_repair / ACV. If >= (threshold - 0.10), flag as PROBABLE TOTAL LOSS. 6. If flagged: a. Append to the CCC/Mitchell/Audatex estimate: - Teardown & Dismantle labor (use shop's blueprint hours matrix) - Administrative Total Loss Processing (2.0 hr flat) - Hazard Storage (per-day, starts at intake) b. Post Slack alert to the shop's estimating channel with: "RO {{ ro_number }} — probable total loss — teardown lines anchored" c. Draft an internal note explaining the flag for the estimator. 7. If not flagged: log the evaluation silently. Take no action. CONSTRAINTS - NEVER write to the estimate without a valid bearer token. - NEVER delete or modify an existing human-authored line. Append only. - NEVER act on confidence < 0.80. Return a claude_code_session_url and set handoff_required = true. - ALL actions must be logged to the audit trail with timestamp, trigger_id, and the specific P-Page or DEG citation used. - If the insurer has invoked a carrier-specific DRP override, respect it and document the conflict in the handoff note.

This prompt is versioned in git. Every change is reviewed. Every shop gets their own variant — labor rates, state thresholds, DRP overrides, and the shop's own teardown hours matrix are injected at routine compile time.

Security, Data Ownership, Reliability

A collision shop handling claims data is a Third-Party Service Provider handling Nonpublic Information under NY DFS Part 500 — whether it knows it or not. Every architectural decision in Collision Intelligence is downstream of that classification. The shop owns the data. Anthropic runs the compute. Intersection Strategies owns the routines. No party sees more than the minimum it needs.

Bearer token authentication

Every routine invocation carries a rotated bearer token. Tokens are scoped to a single shop and a single routine. Revocation is instant. No shared secrets. No passwords in code.

NY DFS Part 500 TPSP posture

The shop is classified as a TPSP handling NPI (claim numbers, VINs, insured names). Collision Intelligence is architected so the shop never exports NPI to an unmanaged destination. Vendor risk diligence is documentable.

Human-in-the-loop via session URL

Every routine that encounters ambiguity returns a claude_code_session_url. A human estimator steps into the live session and completes the reasoning. No silent autonomous action on ambiguous cases.

Append-only audit logging

Every action — every line written, every rental extended, every alert posted — is logged with a timestamp, a trigger ID, the specific P-Page or DEG citation used, and the routine version. Immutable. Exportable for a NAIC audit.

Rate limiting and circuit breakers

Per-routine, per-shop rate limits prevent a runaway cron from spamming the CCC workfile or the rental portal. Circuit breakers trip on anomalous output volume and hand the operation back to a human before damage is done.

Read-only defaults, explicit write grants

Every new routine starts in shadow mode — it reads, scores, drafts, but does not write. Write access to CCC, Mitchell, Audatex, and the rental portals is granted only after the shop operator approves the routine's draft output for a defined period.

Data ownership summary

The shop owns every byte of claim data that passes through the system. Anthropic's enterprise terms prohibit training on customer payloads. Intersection Strategies retains only the minimum metadata required to operate and audit the routines — trigger IDs, timestamps, confidence scores, actions taken. On contract termination, all shop-specific routine variants and stored context are purged within 30 days.

Why Claude, why now

EARLY 2026

Routines shipped

Anthropic's Routines runtime — the asynchronous, trigger-based agent substrate this catalog is built on — shipped to general availability in early 2026. Before Routines, every agent had to be a chatbot. There was no native way to express "fire on a webhook, run 24/7, hand off to a human on low confidence." That is now a first-class primitive.

Q1 2026

Enterprise migration

The enterprise AI landscape has materially shifted. Fortune 500 workloads that ran on GPT-class models in 2024-2025 are migrating to Claude on the basis of tool-use reliability, longer context windows, and Anthropic's enterprise data posture. Collision Intelligence is built where the traffic is going, not where it was.

NOW

The vertical is open

Collision repair has not been won. CCC, Mitchell, and Solera ship estimating platforms; nobody ships trigger-based reflexes that sit on top of them. The first vendor to put a disciplined, audited, trigger-based agent in a shop operator's workflow — with real integrations and real audit logs — owns the category.

The bet

The shop that adopts Collision Intelligence in 2026 is not paying for a model. They are paying for every trigger that has been wired, every DEG citation that has been indexed, every carrier quirk that has been catalogued. The model underneath gets better for free. The moat is the wiring.