Live betaAI Observability

Improve your agents with production feedback.

Trace agent runs, find failures, and test changes to prompts and skills. Your team reviews the results and approves updates before release.

Illustrative exampleAwaiting approval
RuntimeGuardrails check tool calls and results during execution.Agent trafficOn releaseProduction runsLive agent runsDetectFailed checkAutomaticImproveProposedskill changeAutomaticEvaluateCurrent andcandidate resultsAutomaticTeam reviewAwaiting approvalYour team
  1. Production runsLive agent runs
  2. DetectFailed checkAutomatic
  3. ImproveProposed skill changeAutomatic
  4. EvaluateCurrent and candidate resultsAutomatic
  5. Team reviewAwaiting approvalYour team
  6. After release, the new version returns to production runs.
  7. RuntimeGuardrails check tool calls and results during execution.

Evaluation passed. The change waits for team review. Nothing is approved or released.

CtrlB detects failures, proposes a change and evaluates it. The team reviews the results before approval and release.

Turn failed runs into tested changes.

CtrlB flags failures, proposes a change, and evaluates it against the current version. Your team decides what to release.

From a failed run to a tested change
  1. 01DetectFind errors and runs that fail your checks.Automatic
  2. 02ImprovePropose a prompt or skill change based on the failure.Automatic
  3. 03EvaluateTest the proposed change against the current version.Automatic
  4. 04ApproveReview the change and results before release.Your team
Runtime
GuardrailsGuardrails check tool calls and results during execution.Check tool callTool runs if allowedCheck result

Explore self-improving agents

Find the cause of an agent failure.

Follow a failed tool call into the application logs and security events. Use the same evidence to decide what needs to change.

Agent map productionIllustrative example
  • Refund agentSelected
  • Support agentContext only
  • Billing agentContext only
  • Reconciliation agentContext only
  • Payments API
  • Model calls
  • CRM
  1. Agent traceFailedRefund tool call failed
  2. Application logRejectedPayments API rejected the request
  3. Security eventExpiredCredential expired

FindingExpired credential blocked the refund.

Related records need a shared identifier captured in both systems.

A failed refund tool call is linked to an API rejection and an expired credential event.

Inspect prompts, responses, token usage, latency, and tool calls in the run.

Check whether a change fixes the failure.

Save a production example as a test. Compare the current and proposed versions against the same criteria.

Explore evaluations
Illustrative example
Production example
Duplicate email
Saved test
Input
Email a customer.
Reproduce
The send call times out after the email is delivered.
Check
The customer receives one email.
Candidate version
Current versionDuplicate email
Candidate versionOne email
A failed production example becomes a saved test. The candidate version is checked against the same criterion.

Apply policies while agents run.

Check tool calls before execution. Block prohibited actions, request approval when a rule requires it, and mask sensitive data in tool results.

Explore guardrails
Illustrative example
  1. Agent selects tool
  2. Check tool and argumentsBefore execution
  3. Tool runs if allowed
  4. Check resultBefore the agent receives it
  5. Agent receives result
  • BlockedTool did not run
  • Awaiting approvalTool has not run
  • AllowedTool runs, then check 2 masks the email before the agent receives the result
A policy checks the call before execution. It can block the call or wait for approval. A second check masks sensitive fields before the agent receives the result.
CtrlB Relay

Give collaborating agents shared context.

Share progress, findings, and handoffs across agents. Each agent uses that context to decide what to do next.

CtrlB Relay shared contextIllustrative example

Shared goalStop duplicate customer emails.

CtrlB RelayShared contextEvent ledger
  • Triage agent
  • Infra agent
  • Eval agent

Agents use shared context to decide what to do next.

Triage, infrastructure and evaluation agents share findings, work claims, handoffs and results through Relay.

Relay keeps the shared context. Agents organize and adapt around the shared goal.

See the Relay protocol

Discuss your agent workflow

Keep agent data in your own object storage.

Store agent, application, and security data together. Use that context to investigate failures and review changes.

  1. Connect supported tracing.
  2. Send agent runs to CtrlB.
  3. Inspect a run and choose what to evaluate.

View deployment options

Where spans goDirect, no gateway
  • OpenTelemetry SDK
  • Langfuse SDK
  • OTel Collector
  • OTLP exporters
CtrlBParquet in your bucket
  • AI Traces view
  • Evaluations and findings
  • MCP server
  • SQL
  • SOC 2 Type II
  • SSO and RBAC
  • Open data formats
  • AWS Marketplace
Integration details: OTLP, your bucket, and your model key
  • No manual mapping. GenAI-convention spans appear as agent, model, and tool calls.
  • Your bucket. Open Parquet in S3, GCS, or Azure Blob. Readable without CtrlB.Deployment
  • MCP tools. Let your coding agent use search_traces, fetch_trace and run_sql.MCP docs
  • Your model key. Run evaluations and investigations on your key. See the cost per run.
AI Observability pricing

Start with the agents you run today.

Tracing, evaluations, investigations, and Guardrails in one plan.

  • Production traces and model cost tracking
  • Scheduled investigations and prompt comparisons
  • Tool policies with Guardrails
  • OpenTelemetry, your own bucket and your own model key

AI Observability

From $499 / month

Platform fee. Usage is quoted separately for your workload.

Discuss your workload

We agree usage costs before you start. Your providers bill model and bucket costs directly.

Need a larger deployment?Enterprise pricing
FAQ

Common questions

Which agents, providers, and frameworks can I connect?

Send traces from OTLP exporters, including Langfuse and the OTel Collector. Spans that use OpenTelemetry GenAI conventions appear as agent, model and tool calls. Guardrails work with Claude Code, the OpenAI Agents SDK, LangChain, Google ADK and FastMCP.

How are agent runs linked to application and security telemetry?

Through identifiers captured on both sides, such as an order ID or payment request ID. Both records need that shared identifier.

Can I use tracing without the other capabilities?

Yes. Agent tracing and cost tracking work on their own. The other capabilities build on the same traces when you turn them on.

How will usage, evaluations, and guardrails be priced?

From $499 a month for the platform, plus usage. Trace volume, evaluation usage and guardrail checks are quoted for your workload before you start. Model and bucket costs are billed by your providers.

Where is my data stored?

Traces, labels, scores and findings are open Parquet files in your S3, GCS or Azure Blob bucket. Query them with SQL or the CtrlB MCP server. The files stay readable without CtrlB.

How do fixes reach production?

Prompt and skill changes are reviewed in your repository and released through your normal deployment. Guard rules are activated in CtrlB. The Self-Improving Agents page has the detail.

Does CtrlB release changes automatically?

CtrlB proposes and evaluates changes. Your team reviews the results and approves updates before release.

Live beta

Bring a failed run to the demo.

See how CtrlB connects the evidence, tests a proposed change, and records the result.

Live beta. From $499/month, plus usage.