Don't just trust an AI agent. Make it prove every step.
Agent K investigates problems in your AI app, backs up every finding with proof, and only acts when the safety rules say it is safe.
Evidence-backed findings. One safe action. Every decision on the record.
Standing on open, honest tools
SigNoz is the operating system for the whole investigation. Everything the agent sees, does and records lives there.
An investigator you can actually check
Agent K is not a chatbot for your logs. It is a workflow that produces evidence, takes one careful action, and watches its own work.
Evidence feed
Every finding comes with proof
Each claim links straight to the traces, logs, metrics or deployment that back it up. No proof, no claim.
SigNoz over MCP
Investigates through SigNoz
Agent K asks SigNoz for evidence through the Model Context Protocol (MCP). Every query is recorded as its own step.
Sample values, shown for illustration.
Self-telemetry
Watches its own work
It records its own cost, tokens, time and queries, so you can audit how it reached an answer.
Loop breaker
Stops repeated queries
Cost watchdog
Caps spend per run
Guardrails
Knows when to stop
A loop breaker halts repeated queries and a cost watchdog caps spend. When either trips, it escalates to a human.
Rules enforced by code, not by prompts
Three simple rules decide what Agent K may claim, what it may do, and what it must record. Software enforces each one.
Law 1
No claim without evidence
Agent K never guesses. Every finding links to the exact data that backs it up.
- Every claim links to a real SigNoz trace, log, metric or deployment.
- A link checker confirms every published link still resolves.
“The v2 prompt template caused the rise in failed answers.”
Law 2
No action without budget
Agent K can do only one thing: undo the last release. And only when every safety check passes.
- The allowlist holds exactly one action: roll back to the last version.
- If any check fails, it acts on nothing and escalates to a human.
All checks pass
- Objective breached
- Cause tied to a deployment
- Confidence above threshold
Any check fails
- No deployment cause found
- Action not on the allowlist
- Escalate to a human
Law 3
No self without telemetry
Every step is recorded, so you can see what Agent K did, how long it took, and what it cost.
- Token counts and estimated cost are recorded for every model call.
- Loop-breaker and cost-watchdog events are visible in the dashboard.
Sample values shown for illustration, not fixed targets.
From alert to auditable report
Ten fixed steps run in order. The model proposes ideas inside some steps, but code decides what happens next.
Alert received
A SigNoz alert fires and reaches Agent K over a webhook.
Alert received
A SigNoz alert fires and reaches Agent K over a webhook.
The sandbox is real, not a promise
Two design choices make the safety story hold up under scrutiny. Neither depends on trusting the model.
An isolated rollback service
Agent K never holds the keys to your infrastructure. A separate service does the one thing it is allowed to do.
- A small deployer service is the only holder of the Docker access.
- Agent K makes one authenticated call that carries no image reference.
- The deployer picks the last known-good version and records a marker.
A code-only policy gate
The decision to act lives in plain code that you can read and test. The model is not in the loop.
- Checks the allowlist, the alert, the cooldown and the confidence.
- Checks the cause and the sandbox before anything runs.
- Denies the action when any single check fails.
Two safe to fix, two safe to refuse
Each incident is seeded on purpose as a controlled test. Agent K approves a rollback twice and refuses it twice, and it explains why each time.
Prompt-regression deployment
Rollback allowedSymptom
A broken prompt template ships in a new version and failed answers rise right after the deploy.
Investigation
Agent K connects the error spike to the deployment marker and shows the failed traces.
Verdict
The cause is a recent deployment and the error budget is breached, so a rollback is allowed.
Retry-storm cost runaway
Rollback allowedSymptom
A config change lowers timeouts, so the app retries model calls and cost per minute climbs fast.
Investigation
Agent K finds the repeated-call pattern and ties the cost breach to the new deployment.
Verdict
A cost budget breach tied to a deployment is enough to justify a rollback, even with few user errors.
Retrieval latency injection
Rollback deniedSymptom
A flag adds delay to the database search step, so overall response time climbs with no new deploy.
Investigation
The trace waterfall points to the retrieval step as the bottleneck, with no deployment cause.
Verdict
There is no deployment to undo, so a rollback would not help. Agent K escalates to a human.
Database pool exhaustion
Rollback deniedSymptom
Too few database connections are configured, so errors rise and logs show the pool is exhausted.
Investigation
Agent K links the pool errors in logs to the failed traces and explains the failure type.
Verdict
Rolling back the app version does not fix a connection limit, so Agent K recommends a human change.
Reported honestly as four out of four on four controlled scenarios, not as general production accuracy.
Straight answers for a careful reviewer
No. Agent K is a workflow that produces evidence and takes one careful action. There is no free-form chat box. The output is a structured report with links you can click.
Exactly one thing: roll back to the previous version. That is the only action on the allowlist, and it runs only when every safety check passes.
Every claim links to the SigNoz data that supports it. A link checker confirms each link resolves against the running SigNoz instance, so claims without proof never ship.
A policy written in plain code makes the allow or deny call, with no model involvement. By design, two of the four test incidents are refused a rollback.
A separate deployer service is the only part that can touch Docker. Agent K just makes one authenticated call that carries no image reference. Its worst case is a single request to a service that does one fixed thing.
A loop breaker stops repeated queries and a cost watchdog caps spend. When either trips, Agent K stops, marks the investigation incomplete, and escalates with the evidence it has so far.
We report results honestly as four out of four on four controlled scenarios, not as general production accuracy. The scenarios are seeded on purpose so the behavior is reproducible.
Traces, metrics, logs, dashboards and alerts, plus the Model Context Protocol server for queries, deployment markers, and trace-to-log correlation. SigNoz is the center of the whole system.
Prove what it knows. Earn the right to act.
Read the code, watch the demo, and follow the build. Agent K shows its work at every step.