Debugging distributed agents in real time

An automation platform earns trust in one place: the moment something goes wrong. When hundreds of agents run concurrently, “what happened?” has to have a fast, honest answer. Here is how we make agent fleets observable.

Every run is a story

Each agent run in Olsa is recorded as an ordered event log — inputs, decisions, tool calls, retries — not a pile of free-text log lines. Because the engine checkpoints at every await, the story is complete even when a node dies mid-sentence.

Time travel beats print statements

The debugger replays any run against the recorded events, so you can step through yesterday’s failure with today’s eyes. No reproduction environment, no “works on my machine”: the production run is the fixture.

Watching the fleet, not the agent

Single-run stories don’t catch systemic drift. Fleet views aggregate the same event structure across hundreds of concurrent runs — retry storms, slow tools, and permission failures surface as shapes, not tickets.

The principle

Observability is a product feature, not an ops afterthought. If your users can’t see what an agent did and why, they will rightly refuse to hand it anything that matters.