Can I share an AgentOps session replay with someone outside the project?

Not directly — replays sit behind the AgentOps dashboard and project membership. For a stakeholder who needs the outcome rather than the trace, publish a summary page and link the session for anyone who has access.

What's the difference between a session replay and a run report?

A replay is the trace: every LLM call, tool call, token and dollar. A report is the outcome: what the agent concluded, and whether it was right. Engineers need the first; almost everyone else needs the second.

What belongs in an agent run summary?

What the agent was asked, what it concluded, which tools it used, what it cost, where it failed or retried, and a link to the replay for anyone who wants the trace.

Why publish cost and latency to non-engineers?

Because agent economics decide whether a workflow ships. A number nobody outside engineering sees is a number that surfaces first in a finance review.

Can the agent publish this itself?

Yes. At the end of a run, render the summary and publish it — one API call, and the run ends with a URL instead of a dashboard nobody opens.

Share an AgentOps session

AgentOps gives you the thing that makes agents debuggable: a full replay of the session. Every LLM call, every tool invocation, tokens, latency, cost, and the exact point where the agent looped three times before giving up.

It is the right tool for working out why a run went wrong. It is the wrong artifact to send the person who asked what the agent found.

Two different questions

"What did the agent do?" — a debugging question, asked by whoever owns the agent, answered by the replay.

"What did the agent produce, and can we trust it?" — asked by everybody else, and not answered by a trace at all. A replay shows forty tool calls; it does not show whether the conclusion was correct.

Handing a stakeholder a session replay is handing them a flight recorder when they asked where the plane landed.

Publish the outcome, link the trace

At the end of a run, render a summary:

import agentops

session = agentops.start_session()
result = run_agent(task)
stats = agentops.end_session("Success")

publish_report(
    title=f"Agent run — {task_name}",
    html=render_summary(
        task=task,
        conclusion=result.summary,
        artifacts=result.outputs,
        cost=stats.cost,
        duration=stats.duration,
        replay_url=session.session_url,
    ),
)

The page carries, in this order: what was asked, what the agent concluded, the evidence, what it cost, and — at the bottom — the AgentOps replay link for whoever has access and wants the trace.

That ordering is the whole design. The reader who needs the answer gets it first; the engineer who needs the trace is one click away.

Cost belongs on the page

Agent runs cost real money, and that cost is usually invisible outside the team running them until somebody reads a bill.

Putting $0.84 · 4m 12s · 37 tool calls on every run summary changes the conversation from "is this agent good?" to "is this agent worth it?" — which is the question that actually decides whether the workflow ships. It also makes regressions visible: a run that cost $0.84 last week and $3.10 today is a problem you find now rather than at month end.

Feedback closes the loop

The reason to publish rather than post is that the reply becomes input. A reviewer who comments "this conclusion ignores the Q2 data" on the relevant paragraph is giving the agent its next instruction — and the agent can read it:

Fetch open comments on report <id>, address them, update the report, and reply in each thread.

See letting an agent respond to comments.

Worth knowing

Do not paste the trace into the report. Forty tool calls rendered inline buries the conclusion; the replay link covers it.
reports:write, plus comments:read for the loop — see scoped tokens for AI agents.
One report per recurring task, updated per run, so the URL tracks the latest outcome with its history attached.
5 MB per report body.

Try it

Free — unlimited reports, commenters and revisions.

Read the API reference →