AI agent observability and reliability for production. Trace execution, diagnose failures, and recover multi-agent workflows from any checkpoint.
/FailuresReal-time diagnostics and anomaly detection for your agent swarm.
- 2m agoSchema guard added to lookup_order74 paused runs resumed · 0 dupes
- 14m agoRetry wrapper on tool.searchFlaky rate dropped from 4.2% → 0.3%
- 1h agoContext dropped in llm.chatSuggest trim(MAX_TOKENS) before plan step
Failures over the last 4 months
Saved vs last week via auto-recovery.
- traces captured
- 94M+traces captured
- failure recall
- 99.2%failure recall
- frameworks instrumented
- 47frameworks instrumented
Everything looks healthy. The agent still produced a broken answer.
Omium/semantic traceOne platform for AI agent observability, failure analysis, and checkpoint recovery. Designed around how production agent runs actually fail — not generic microservice health checks.
Execution checkpointingSnapshot agent state at every tool call, LLM response.Restore from any moment.
- search_docs()tool
- llm.call(gpt-5)llm
- branch: needs_lookupbranch
- db.query(orders)tool
- llm.call(gpt-5)llm
- return_response()tool
Failure classificationEvery silent failure auto-tagged: hallucination, infinite loop, tool error, context drop.
- r_9af2supportAgent3:44schema-miss
- r_9ad7ingestAgent3:41hallucination
- r_9ac1supportAgent3:38tool-error
- r_9ab6plannerAgent3:36infinite-loop
- r_9a9esupportAgent3:31context-drop
Fix suggestionsAI analyzes failure traces and proposes code-level fixes — not just alerts.
async function lookupOrder(id) {- const res = await db.query(id)- return res.rows[0]+ const res = await db.query(id)+ if (!res.rows.length) throw NotFound+ return res.rows[0]}
Checkpoint recoveryResume from any checkpoint. Re-run with new inputs. Fork timelines.
Real-time tracingStructured spans for every LLM call, tool use, and agent decision. No sampling.
Time-travel replayRewind any production run. Inspect inputs, outputs, reasoning. Deterministic.
No rewrite. No sidecar. Drop the SDK into any agent framework and start getting traces within the minute. From install to checkpoint recovery, the four steps below take you from unobserved production to full AI agent observability.
- 01
Install
Install the Omium SDK with one line. TypeScript, Python, or Go clients share the same tracing primitives, so polyglot agent stacks instrument the same way. No collectors, no sidecars, no rewrites — point the SDK at your workspace and your first production AI agents start reporting within the minute.
- 02
Instrument
Wrap your agent code. Zero config required — Omium auto-detects LangChain, LangGraph, OpenAI, Anthropic, and custom agent loops and captures their tool calls, prompts, schemas, and retries without hand-rolled spans. Multi-agent systems get a single unified trace spanning every hop so cross-agent failures stop hiding between services.
- 03
Observe
Agent traces stream live into the dashboard with sub-second ingest. Filter by run ID, user, model, or failure class; search across every tool call and LLM message; alert on schema drift, tool-call loops, budget blowouts, or silent regressions. Reliability engineers get production visibility without custom dashboards.
- 04
Recover
When a run fails, Omium diagnoses the root cause and lets you resume from any checkpoint — not just restart from zero. Schema guard catches output mismatches; failure analysis ranks likely fixes; paused-run recovery replays an agent from the last good state so production workflows survive bad completions, rate limits, and upstream outages.
Every layer of your agent's execution, inspectable. Hover a disc to isolate the layer.
- 01
Live alerts
Active failures & anomalies
- 02
Metrics
Latency, error rate, token cost
- 03
Traces
Structured spans, causal chains
- 04
Raw event stream
Unindexed logs & tool calls
Structured spans for every LLM call, tool use, and agent decision. Search across millions of traces. Drill down to the exact moment things went wrong.
When a run fails, Omium's analyzer diffs it against thousands of successful runs and proposes exactly what changed — with a suggested patch.
Tool lookup_order returned an unexpected payload shape on prod/us-east-1. Downstream response-writer would have hallucinated billing totals. Run paused at ckpt_0a4c before any side effects were committed.
Omium Fixlookup_order returned a mismatched shape.response-writer.Zod guard on lookup_order output, mapped the renamed field total_cents → amount, and re-ran ckpt_0a4c for all 74 paused runs. Passing new schema to response-writer.Branch: omium/AGT-1487-schema-guard-9214
Every agent run is checkpointed at every critical step. When something breaks, rewind to the last good state and resume — or fork a new branch with different inputs.
supportAgent·run_8f2aOn it. Tracing live execution of supportAgent.
Captured checkpoint ckpt_0a4c · tool_callbefore tool invocation.
Running lookup_order in prod/us-east-1.
Tool returned non-JSON payload. Schema validation failed.
Pausing before downstream call. Snapshot captured: ckpt_0a4d · pre_llm
Drop-in support for LangChain, LlamaIndex, CrewAI, AutoGen, OpenAI Assistants, Anthropic MCP, and more. Auto-instrumented — no code changes.
LangChain
LlamaIndex
CrewAI
LangGraph
Pydantic AI
Mastra
Vercel AI
OpenAI
AnthropicLangChainLlamaIndexCrewAILangGraphPydantic AIMastraVercel AIOpenAIAnthropic
Google
Mistral
Cohere
BAML
Guardrails
AutoGen
HaystackGoogleMistralCohereBAMLGuardrailsAutoGenHaystackThe Omium SDK wraps any agent function without config. It auto-detects common frameworks and instruments them at runtime.
14,829 agents tracked now
No credit card to start. Upgrade when you need higher limits, longer history, or enterprise compliance.
Free
For personal projects and small experiments.
Billed monthly
Get Started- 500 agent runs / month
- Core observability
- 7-day data retention
- 1 seat
Developer
For solo developers shipping agents in production.
Billed monthly
Start trial7-day free trial
- 2,500 agent runs / month
- Core observability
- Search + filters
- User feedback
- Local-first mode
- 14-day data retention
- 1 seat
Pro
For teams running scale-out workloads with higher retention.
Billed monthly
Start trial7-day free trial
Everything in Developer, plus:
- 25,000 agent runs / month
- Silent failure detection
- Topic clustering
- 60-day data retention
- Up to 5 seats
- Slack support
Enterprise
For regulated deployments and dedicated infrastructure.
Everything in Pro, plus:
- Unlimited seats
- Custom rate limits
- Custom data retention
- SSO login
- SLA guarantees
- Prioritized support