Understand what hundreds of agents are doing
Agents at work all day, across every model you use. The hard part is no longer reading one agent's transcript. It is knowing what the whole operation is doing, what backs that picture, what you do not know yet, and where to steer next.
Your analysis agents turn logs, transcripts and other records you allow them to read into a small set of living pages: goals, streams of work, what actually happened, open questions and what is still unread. The next agent continues from those pages instead of writing yet another report.
Oddlynew does not look inside models, detect misbehavior on its own, or stop agents. Those jobs belong to the systems that run your agents. Oddlynew holds the shared picture and the decisions you make from it.
Who this is for
teams running many agents for hours or days;
AI labs and evaluation teams studying how agents behave together;
software teams with many agents working at once;
safety and incident teams looking into what agents did;
anyone whose agents now do more than one person can follow.
The problem
A large agent operation produces more than any person or single model can hold in mind. Adding analysis agents helps, but on its own it leaves you with hundreds of separate summaries, different words for the same thing, gaps nobody sees, and confidence nobody can check. What helps is one picture that many agents keep up to date.
What to keep
Organize around the work, not around the agents. One page per agent is rarely useful. Pages that tend to earn their place:
Current picture: the best current summary of what is happening and why.
Goals and workstreams: what the operation is trying to do, and the streams of work that serve it.
What actually happened: actions, system changes and receipts recorded outside the agents themselves.
Claims and evidence: each important conclusion, linked to what supports it.
Competing explanations: alternatives that stay visible until they are settled.
Who directs whom: which agents or workstreams steer others.
Access: the tools, permissions and environments the operation can use.
Coverage: what has been read, what has not, and which data may be missing or unreliable.
Steering decisions: changes to goals, scope or access, with the receipt from the system that carried them out.
Add structure only when the same need keeps coming back.
Several analysis agents, one picture
Give different agents different jobs and let them all update the same pages. For example, one maps goals and workstreams, one rebuilds the timeline, one traces who directs whom, one tests other explanations, one checks what has not been read yet, and one challenges the current summary. Any capable agent can take any of these jobs; what matters is that their conclusions end up in the same place.
Keep different kinds of truth apart
At this scale a polished summary can be more dangerous than open uncertainty. Keep these apart:
What happened: recorded independently of the agents.
What an agent says: what it claims it intended, believed or did.
What we think: the current explanation from an analyst, human or agent.
What does not fit: facts that contradict the current picture.
What we have not seen: activity that is missing, unread or inaccessible.
When new evidence changes the story, update the picture and keep the earlier explanation and the reason it changed.
Steer from the shared picture
You do not need to approve every action to stay in control. From the shared picture you can change the goal or scope, focus on one workstream, ask for a counter-analysis, challenge a conclusion, remove an access, isolate a workstream, or ask the system that runs the agents to pause or stop them.
Oddlynew records the decision, the evidence behind it and the receipt that comes back. Whether a pause or stop actually happened is decided by the system that runs the agents.
Where Lenses fit
A Lens shows the pages that matter right now in one view, and lets you see the work as it happens and correct its course. At fleet scale that is a map of goals, workstreams, claims, evidence and gaps, not one card per transcript. A clear Lens does not prove the picture is complete; coverage and open questions stay visible in it.
A good first task
Start with the records of one multi-agent run that you are allowed to share. Ask your agent:
Read this organization's AGENTS.md and the agent-run records I give you. Build a current picture of the run. Keep what happened, what agents said, what you think, what does not fit and what you have not seen apart. Organize around goals and workstreams, not one page per transcript. If you use more agents, have them update the same pages. Tell me what we believe, what supports it, what we may be missing, and which next step would reduce the uncertainty most.
Then give a second agent the same organization and ask it to challenge the picture. It should be able to continue from the pages instead of starting from the raw records.
What good looks like
A person understands the operation without reading every transcript.
Important claims link to their evidence.
What happened and what agents said are never mixed up.
Uncertainty and gaps stay visible.
Several agents improve the same picture.
Someone else can pick up the investigation where it stands.
Steering decisions and their results come back as records.
Good to know
Oddlynew does not provide model interpretability, secure telemetry, anomaly detection, sandboxing, access revocation or an emergency stop. Those belong to the model, the runtime, your infrastructure and your security systems. A convincing picture built on incomplete records is still wrong, so keep sources, coverage and competing explanations in view.
Why it gets better
Every investigation leaves more behind than a report: shared words, known sources, a clearer map of the work, better methods and known failure modes. The next agent starts from the current picture, its evidence, its open disagreements and the places where understanding is weakest.