# Build agents that learn from their mistakes.

CtrlB finds what went wrong in production, proposes a prompt fix, and tests it against the runs that failed. Your team decides what ships. Self-improving agents, traces and costs, evals and Guard are all live.

## From a failed run to a tested fix

Scheduled investigations group failed runs by cause and suggest prompt changes, with traces as evidence. Save the runs as a test set, compare the old and new prompts on the same cases, and review the results before shipping a change.

Agents send OpenTelemetry spans (OTel GenAI conventions, Langfuse and other OTLP exporters) to CtrlB. Traces, scores and findings live as open Parquet in your own S3, GCS or Azure Blob bucket and remain readable without CtrlB.

## Capabilities

| Capability | Status | Detail |
|------------|--------|--------|
| Agent traces | Live | Complete agent runs, model calls and tool calls. Readable through the CtrlB MCP server (`search_traces`, `fetch_trace`, `run_sql`) |
| Costs | Live | Model, tokens in and out, latency, cost and prompt hash per call; tool name, arguments, result and status |
| Evals | Live | A judge model, on your own provider key, scores production traces; scores are stored next to the trace |
| Self-improving agents | Live | Investigations group failed runs by cause, produce findings with a suggested prompt fix, and replay frozen trace sets against the new prompt |
| Guard | Live | Checks each tool call in the agent framework's hook and answers allow, deny, ask or rewrite; every check is written to your lake |

## Metadata

- **Canonical HTML**: https://ctrlb.ai/ai-observability
- **Last updated**: 2026-08-26

## Pricing

- AI Observability starts at $499/month for the platform; usage is quoted separately
- Includes tracing, evaluations, scheduled investigations, prompt comparisons and Guard
- Trace volume, evaluation usage and Guard checks are agreed before starting
- Model and bucket costs are billed directly by the customer’s providers
- Enterprise deployments have custom pricing
- Evals and investigations run on your own model key; your provider bills the judge's tokens

## Links

- [AI observability pricing](https://ctrlb.ai/ai-observability#pricing)
- [Book a demo](https://ctrlb.ai/contact)
- [Documentation](https://docs.ctrlb.ai/)