An agent’s actions
need a record.
We’re building toward infrastructure that connects what a system was asked to do with what it actually did and what happened next.
The task is only the beginning.
Once an agent can take actions, the result alone is an incomplete account of its work.
A completed task may involve new permissions, a changed environment, an unexpected external system or a human intervention. Those details determine whether another run is comparable and whether the same behavior should be allowed again.
We want to preserve the relationship between the objective, the instructions, the available tools, the action sequence and the outcome. An agent’s own explanation is one piece of evidence, not direct access to its reasoning or intent.
A publication. A record method.
A software prototype.
| Work | What it contributes | Current state |
|---|---|---|
| Dispatches | Concise reporting tied to primary sources. | Initial editorial collection |
| Observation records | A proposed structure for behavior, context and evidence. | One worked example |
| Recorder | Experimental tooling for recording events and checking specified behavior. | Prototype; validation work remains |
Recorder is an initial building block. Its current checks do not establish broad agent reliability or coverage of unobserved actions. A useful result needs to say which behavior was examined, which evidence was available and where the check does not apply.
Inspect the observation format →Follow the event.
Keep what can be learned.
Dispatches surface developments worth understanding. Observation records preserve the supporting evidence and its limits. Recurring questions across those records guide the software we build.
Over time, a useful corpus could connect incidents, successful runs, near misses and interventions across versions and environments. Comparisons would retain the conditions under which behavior occurred, rather than flattening everything into a model score.
This is the direction of the work. We do not yet operate a cross-provider telemetry network or a production monitoring service.
Earn the comparison.
The next step is to test the record format against traces from real deployments: incomplete logs, changing permissions, human handoffs and outcomes that take time to become visible.
That work should establish what can be captured reliably, what must remain unknown, and which recurring behaviors are suitable for reproducible checks. Monitoring and intervention should grow from that evidence.
For teams deploying agents, the starting question is concrete: can you reconstruct a consequential action well enough to decide what should happen the next time?
Building agents that act?
We’re interested in the moments when the task, the permissions and the actual behavior stop lining up.
Talk to us about an observation, a deployment or a technical collaboration.
contact@agentobservatory.dev