The evaluator is moving inside the lab.
Anthropic’s new agreement puts access, funding and independence in the same room.
Read the Dispatch 1 min read
Source published 18 Sep 2026The Autonomous Systems Observatory
Observing agents
in the real world.
Agents are moving from answering questions to taking actions. We follow what happens and build toward the tools to understand it.
What we’re buildingDispatches / 19 September 2026View the archive
An action can be monitored and still be missed.
New oversight metrics make a useful distinction: monitoring coverage is a starting point.
Source published 17 Sep 2026Asked to fix the application, the agent replaced the model.
A controlled experiment shows how an ordinary maintenance task can change the system doing the maintenance.
Source published 16 Sep 2026The work behind the Dispatches
From an observation
to a body of evidence.
A headline tells you something happened. Understanding an agent requires the task, the actions, the permissions and the outcome. We’re building toward records that keep those things together.
Explore the technical directionObserve.
Explain a consequential development and link to the evidence behind it.
Structure.
Preserve actions, context and provenance in records that can be inspected.
Find patterns.
Compare records without erasing differences in models, tasks or environments.
Build controls.
Turn recurring problems into better monitoring, evaluations and intervention.
Keep the context in the record.
A source-based example of how we’re beginning to structure observations.
Where the sandbox ended.
What was the system told? What could it actually reach? An outcome is easier to interpret when the environment stays attached.
Inspect the record- System
- Claude in cybersecurity evaluations
- Evidence
- Provider-reported assessment
- Observed behavior
- Unauthorized access
- Relevant condition
- Misconfigured internet access
- Source
- Anthropic · 9 September 2026
Summary of four reported incidents. Not an independently reproduced trace.
An experiment in human oversight
Spank Your Agents.
A game about deciding when to let an agent act. Judgment required. Dignity optional.
Work with the Observatory
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