The 8 AI agents behind Car Diagnostics: how a specialized AI team explains your car fault

A classic AI assistant is one big chat (GPT-4, Claude, Gemini) you ask questions to. The problem for accurate car diagnostics: one general model isn’t enough — you need expert knowledge of DTC codes, of make/model specifics, of a car’s own history, of live telemetry, of failure prediction. That’s why Car Diagnostics runs eight specialized AI agents, each with its own role. This post covers what each does and how they work together.

Why 8 agents and not one

A single large agent has to “know everything” — and ends up knowing everything mediocrely. Specialized agents:

• Have narrow, deep context (say, only Toyota DTC codes)
• Are significantly faster (smaller system prompt, fewer tokens)
• Are more accurate (specialized prompts, specialized training sets)
• Are cheaper to run (small models for simple tasks, large ones only where needed)
• Can run in parallel (multi-agent orchestration)

Our stack is built on the principle that every agent has its own narrow DB access, its own memory section, its own MCP endpoint.

1. OBD Reader Agent — decodes raw adapter bytes

Role: takes raw bytes from the ELM327 (e.g. 41 05 7B) and turns them into human values: Coolant Temp: 83°C.

What it does:

• Decodes 200+ standard SAE J1979 PIDs
• Handles ISO-TP multi-frame (VIN, DTC lists)
• Recognises adapter quirks (ELM327 v1.5 vs v2.1, clones)
• Auto-detects the protocol (ISO 15765-4 CAN, KWP, PWM, VPW)

When it runs: on every session — it’s the “translator” between the adapter and the rest of the system.

2. DTC Analyzer Agent — interprets fault codes

Role: takes a DTC (P0420) + freeze frame + session history → returns code description + likely causes + a diagnostic plan.

What it knows:

• Catalog of 1947+ DTCs (SAE J2012, ISO 15031-6) — see the DTC catalog
• Make-specific patterns (for Toyota P0420 is more often an exhaust leak, for VW — the O2 sensor)
• Code interrelationships (P0300 + P0171 = almost always a vacuum leak)
• Current parts pricing (to estimate repair cost)

When it runs: after every DTC scan — see the P0420 diagnostics post.

3. Prediction Agent — forecasts failures before they happen

Role: analyses session history (LTFT, catalyst efficiency, O2 oscillation) and predicts the probability of a component failing in the next 300-3000 miles.

What it does:

• Compares your numbers against 100k+ cars of the same make/model/mileage
• Detects “trends” — not one-off anomalies but stable drifts
• Estimates remaining life for O2 sensor, MAF, catalyst, injectors
• Warns 2-3 weeks before Check Engine actually appears — so you can order the part in advance

When it runs: weekly, after 3+ sessions. Notifies in the app.

4. Recommendation Agent — gives step-by-step advice

Role: after diagnosis (DTC + LTFT + freeze frame) issues concrete steps for the driver: “First clean the MAF sensor; if it doesn’t help — run a smoke test for a vacuum leak.”

What it considers:

• Your DIY level (basic, intermediate, advanced — from profile)
• Available tooling (multimeter, compression tester, smoke generator)
• Weather (EVAP diagnosis doesn’t work in freezing temperatures)
• Cost — from cheapest actions to most expensive

When it runs: right after DTC Analyzer, complementing its result.

5. Report Generator Agent — produces PDF reports

Role: aggregates the full session (DTC, freeze frame, live parameters, history) into a readable PDF report for a shop, insurance company or used-car buyer.

What it does:

• PDF layout with LTFT/O2 graphs, DTC tables, DPF/catalyst imagery
• Localization (EN, UK, soon RU, PL)
• “Mechanic mode” (protocol numbers) or “Owner mode” (plain explanations)
• Ready-made template to hand a car over at sale

When it runs: on user request — “Generate report” button on the session screen.

6. Chat Agent — conversational assistant

Role: when you have a question like “my car dies at the stop light — what to do?” — Chat Agent runs the dialogue, asks clarifying questions, pulls in previous session results.

What it does:

• Remembers previous conversations (per-plan chain state — see the How the AI works page)
• Pulls in other agents’ results (DTC Analyzer, Prediction) into the dialogue context
• Directs you to the right app screen (“open Live Monitor and read LTFT”)
• Understands non-technical descriptions (“rattles”, “jerks”, “smells like fuel”)

When it runs: whenever you write in the app chat.

7. RAG Agent — knowledge base access

Role: when another agent needs specifics — say, “how exactly do I test the EGR valve on a Ford EcoBoost 1.5” — it queries the RAG Agent (Retrieval-Augmented Generation). That agent pulls the exact answer from a knowledge base built on manufacturer docs, forum threads, service bulletins.

Knowledge base contains:

• All OEM service bulletins for the most common models
• Forum threads from 20+ specialized communities
• Wikipedia + technical handbooks
• Our own content — see the blog and docs

RAG architecture — a vector database with text embeddings retrieves relevant technical material and injects it into the model context. See How the AI works for the full technical description.

8. Diagnostic MCP Agent — orchestrator

Role: the main dispatcher — receives a request and decides which of the other agents to involve. For example:

• Simple question “what does P0420 mean?” → straight to DTC Analyzer
• Complex “my car dies at the light, what should I do?” → Chat Agent + DTC Analyzer + Recommendation
• PDF report request → Report Generator + all preceding session results

Diagnostic MCP also handles rate-limiting, billing (expensive-model agents are paid-tier only) and shared session memory.

How to see it in action

Try it on the free tier — you get OBD Reader, DTC Analyzer, Chat Agent (limited to 10 messages/day) and basic Recommendation. Prediction, Report Generator and full Chat are on the paid tier.

Next time you see Check Engine — pair an ELM327 adapter, scan the car, and watch the 8 agents work together: in 30 seconds you get a code explanation, likely causes, step-by-step actions, a failure forecast and optionally a PDF report for the shop. That’s what no other OBD app does today (deep comparison — see Car Diagnostics vs Torque Pro).