Agents ▸ Invocations
Every agent run is recorded as an invocation, whether it succeeded, failed or is still waiting on a person. This is the run log, the screen to open when an agent did not do what you expected, or when you want to know what something cost.
The list
Filter by Agent, Provider / model, a From and To date, and Termination. Columns show the agent, provider and model, when it started, cost in USD, how many recommendations it produced, and a termination pill. Sort on any of them.
Sorting by cost is the quickest way to find an agent that is more expensive than it is worth. Filtering by termination is the quickest way to find a systemic failure.
The detail view
The header strip carries the run's facts: agent, provider and model, start time, duration, tokens in and out, cost, recommendation count and termination reason. A run created by replaying an earlier one is badged Replay of #N.
Four tabs:
Summary: the termination reason with its remediation, and the agent's Decision summary in its own words. Start here.
Recommendations: what this run proposed, read-only, with a Review in inbox link to act on them.
Structured result: the machine-readable result the agent emitted, as JSON. This is what later actions in the same hook can read. Useful when a downstream action behaved oddly.
Transcript: the whole run, step by step: the opening prompt, each tool call with its arguments and response, the agent's reasoning between steps, and the closing output. Failed steps open automatically; everything else can be expanded or collapsed.
The transcript is where you find out why an agent decided something. If an agent's output is wrong, it is nearly always visible here as either a tool returning less data than you assumed, or the agent reasoning correctly from a bad premise in its prompt.
Termination reasons
| Reason | What it means and what to do |
|---|---|
| Completed | The run finished normally. |
| Error | The invocation failed. Check the transcript and the batch action output. |
| Iteration cap reached | The agent used its full tool-call budget before finishing. Narrow the input, or raise the agent's maximum iterations. |
| Waiting for a human response | The agent escalated. The job resumes when the person responds. See Escalations. |
| Monthly cost cap reached | Month-to-date agent cost hit the cap. Raise or remove it under Agents ▸ Configuration. |
| API key missing | No provider key is configured for this client. Add one under Providers & budgets. |
| API key unusable | The provider rejected the key, or it could not be read from storage. Re-enter it under Providers & budgets. |
| API key disabled | The key is switched off. Enable it under Providers & budgets. |
| Insufficient credit | Your provider account is out of credit. Top it up. |
| Rate limited | The provider rate-limited your key. Wait, then re-run, or check your account's limits. |
| Model tier unresolvable | The agent's tier has no matching model for the configured provider. Check the provider on the install. |
| Library agent unavailable | The installed agent is not in the current release. It works again if a release restores it; otherwise point the action at another agent. |
| Output cut off | The agent hit its per-turn output limit mid-answer, so the result is incomplete. Raise the per-turn output budget, or ask it to produce less at once. |
Every one of these is recoverable, and none loses the triggering job: fix the cause and re-run.
Re-running
Hook batches containing an agent action offer two ways to re-run, in the Re-run agent batch dialog:
- Replay (free): re-presents the recorded result. No provider call, no cost, no side effects.
- Fresh run: re-invokes the agent against current data. This costs money.
Replay is for reproducing what happened, checking a downstream action against a known agent result. It cannot give a different answer, because it does not ask the agent anything.
Fresh run is for after you have changed something: fixed the key, raised the cap, edited the prompt, corrected the content. It is also the only option that reflects data changed since the original run.
Both are available as mass actions on the batches list, as Rerun (replay agent result) and Rerun (fresh agent run).
Costs
Cost per invocation is estimated from recorded token usage and published model prices, and will not match your provider's bill exactly. The Providers & budgets cost meter aggregates the same figures.
The transcript explains an expensive run: agents cost roughly in proportion to how much content they read, so an audit agent over a long page, or one that made many tool calls before finishing, will show it there.