Know what every agent in the crew is doing.
A crew of agents delegating to each other is powerful — and completely opaque when something goes wrong. Obsivara traces every agent's tasks, every delegation, every tool call, and every model invocation, so you can see the crew work instead of guessing.
CrewAI, fully observable.
Per-agent task tracing
Each agent's tasks traced end to end: the instructions it received, the tools it used, the output it produced, and how long it took.
Delegation flow visibility
See the full chain of agent-to-agent delegation, where handoffs added value, and where they just added tokens.
Crew-level cost breakdown
Cost attributed per agent, per task, and per crew run — find out which agent burns 70% of the budget and whether it earns it.
Failure and retry analysis
Failed tasks, retries, and tool errors classified and grouped, so recurring crew failure patterns get fixed once instead of debugged repeatedly.
Three steps. No code changes.
Add the Obsivara instrumentation
Drop Obsivara's tracer into your CrewAI project — a few lines of setup, no changes to agents, tasks, or crews.
Crews report in automatically
Every crew run streams to Obsivara with agent-level traces, delegation maps, and cost.
Set crew budgets and alerts
Per-crew cost caps, task-duration limits, and failure-rate alerts keep autonomous crews inside guardrails.
Signals tracked out of the box.
- →Task duration and status per agent
- →Delegation chains between agents
- →Tool invocations and error rates
- →Tokens and cost per agent and task
- →Retry loops and stuck tasks
- →Crew run success rate over time
Common questions.
Does Obsivara change how my crews execute?
No. Instrumentation is observe-only — Obsivara records what happens; it never alters delegation, task order, or model calls.
Can I compare different crew configurations?
Yes. Run variants side by side and compare cost, duration, and success rate per configuration on real workloads.
What happens when a crew gets stuck?
Obsivara flags stalled tasks and abnormal retry loops in real time, with the full trace showing exactly which agent and which step stalled.