Autopilot

Maev Autopilot is the active intervention layer that proxies, analyzes, and optimizes your agent's LLM calls. It utilizes a Hybrid Architecture by coupling our native SDK with a self-improving Gateway.

How the Hybrid Architecture Works

To get absolute control over your agent's execution loop (like local circuit breakers and local tool interceptions), you install the maev-sdk directly into your codebase. When you execute maev.run(agent, gateway=True), it achieves two things simultaneously:

1. Client-side Interception: It patches your Python agent loop so Maev can natively intervene if an infinite tool loop occurs or a budget circuit breaker is hit.
2. Transparent Proxying: It rewrites the base URL on your LLM client instances to point traffic through https://gateway.maev.dev/v1 — an OpenAI-compatible endpoint that routes to any provider.

Level 1: Immediate Protection

Because the gateway supports all major providers natively, you instantly get:

• Provider Fallbacks: If Anthropic fails, seamlessly route to OpenAI without crashing your agent.
• Auto-Retries: Native retry logic on rate limits or random API drops.
• Full Observability: Every request is captured with its exact latency, cost, and token usage.

Level 2: Automated Self-Reflection (After enough failures)

Once an agent accumulates enough failures, Maev's backend triggers an offline optimization cycle.

1. Failure detected via the rules engine.
2. The failing prompt and output are analyzed by Maev's backend.
3. A new prompt candidate with targeted adjustments is generated automatically.
4. The candidate is evaluated offline against past successful runs before being tested.

Level 3: Shadow & Canary Traffic

Once a candidate passes offline evaluation, the Gateway tests it in the real world:

• Shadow Mode: Incoming traffic is duplicated and run against the new candidate. The response is compared to the original, but never returned to the user.
• Canary Promotion: If shadow mode performs well, a small percentage of real traffic is routed to the new candidate.
• Live Promotion: If canary succeeds with higher quality and lower error rates, the new candidate replaces the base prompt entirely.

Observing the Process

Open any session in the Maev dashboard and look for:

• Optimization Activity — See exactly what shadow candidates are currently running in the background.
• Promotions & Rollbacks — Track when your agent gets formally upgraded to a newly generated prompt.