Autopilot

The Autopilot tab shows how Maev is protecting and improving your agents over time.

Stats cards

At the top of the page, four cards summarize Autopilot activity across all your agents:

Run list

Below the stats, each row is one agent run. You can see:

• Agent name and run ID
• Status: success, failed, circuit_broken, timeout
• Cost, duration, and LLM call count for the run
• Number of Autopilot interventions (retries, fallbacks, circuit breaks) that fired

Click any row to see the full session timeline with every LLM call and intervention logged in order.

Strategies

The Strategies section lists every learned recovery pattern Maev has identified for your agents. Each strategy shows:

• The failure pattern it addresses (e.g. empty response on a specific model, rate limit on a specific prompt type)
• The recovery action (temperature nudge, model switch, retry delay)
• Win rate and number of times applied
• A plain English explanation of why it works, generated automatically

Strategies are scoped to the agent, to your organization, or globally (Teams plan). You can disable any strategy or delete it.

Autopilot protections

Every maev.run(..., gateway=True) call activates these protections automatically:

Auto-retry

If an LLM call fails (rate limit, timeout, empty response), Maev retries it up to 3x with exponential backoff. If the same prompt returns an empty response twice in a row, Maev nudges the temperature slightly before retrying.

No configuration required. Works across OpenAI, Anthropic, LangChain, LiteLLM, and Gemini.

Cost circuit breaker

If the accumulated cost of LLM calls in a single run exceeds cost_budget (default $1.00), Maev stops the run immediately and raises CircuitBrokenError. The partial run is saved to your dashboard.

maev.run(agent, gateway=True, cost_budget=0.25)  # stop if this run exceeds $0.25

Call count limit

If more than max_calls LLM calls fire in a single run (default 50), Maev stops the run. This catches infinite loops before they drain your budget.

maev.run(agent, gateway=True, max_calls=20)

Duration limit

If the run takes longer than max_duration_s seconds (default 300), Maev stops it.

maev.run(agent, gateway=True, max_duration_s=60.0)

Loop detection

Maev hashes the last 20 prompts in a run. If the same prompt hash appears 3 times, the run is stopped as a detected loop. This catches cases where an agent is stuck in a semantic cycle even if the exact text varies slightly.

Fallback models

If the primary model fails repeatedly, Maev automatically switches to the next model in your fallback list.

maev.run(
    agent,
    gateway=True,
    fallback_models=["gpt-4o-mini", "gpt-3.5-turbo"]
)

How Autopilot learns

Autopilot operates at three levels:

Level 1: Immediate (every run)

Mechanical protections that fire instantly: retries, circuit breakers, fallback models, loop detection. These activate from run 1 with no data required.

Level 2: Pattern learning (after 20+ runs)

After 20 runs for an agent, Maev starts analyzing patterns:

• Which retry strategies recovered the most failures
• Which models produce the fewest empty responses for this agent's prompts
• What temperature ranges produce the best outputs
• Which intervention sequences worked vs did not

Winning patterns become strategies that are applied automatically in future runs.

Level 3: Cross-agent strategies (Teams plan)

Strategies that prove effective across many different agents in your organization become global strategies. These apply to new agents before they have accumulated enough runs to have their own patterns.

Viewing interventions

Every Autopilot action is logged to the session timeline. In the Sessions detail view, look for intervention events:

• retry — an LLM call was retried and the reason
• fallback — switched to a fallback model
• circuit_break — run was stopped and why
• loop_detected — prompt loop detected
• strategy_applied — a learned strategy modified a call