Agent Runs

Agent Runs is the live operations log and human-in-the-loop inbox for every job handed to your Agent Team. Watch runs stream step-by-step in real time, filter by roles or runs waiting on your input, answer agent questions and permission cards inline, inspect every tool call and screenshot, and jump straight to the Proposed Changes a run filed.

Key capabilities

• Status & role filtering — filter runs by Open (queued, running, needs input), Needs you (needs input), Done (completed), Failed (failed, cancelled), or All, and narrow by any role on the team
• Real-time streaming transcript — live Server-Sent Events (SSE) stream with automatic 5-second polling fallback, showing agent reasoning (Thinking), Markdown messages, connected tool-call traces with duration and preview, screenshots, and the final Result summary
• Inline human-in-the-loop cards — when an agent needs a decision, the run parks in Needs you (needs input) and renders an interactive card directly in the transcript: Permission (Allow, Allow always, Deny), Question (one-click option buttons or free-text reply), Sign in (Open sign-in page + I signed in / Skip), or Secret (one-time password input, never stored on the row)
• Calls to connected services, named — when an agent uses a server you connected under Integrations, the transcript shows the service and the tool it called (for example Gusto · get_payroll_register) with an External tag, so you can tell your own books from outside data at a glance
• Proposed changes, decided where the job is — every gated write a run proposes appears inline in the transcript (Proposed pill) and under Proposed changes in the run drawer and in the job's Team chat post, each with its tier, status and amount, Approve / Reject on waiting ones (a rejection needs a reason) and Open in Books to the full change in Proposed Changes, where the card names the teammate that filed it and links back to the job
• Every job in Team chat — each run is a post in the business's shared Team chat with the job's full Activity log (reasoning, steps, replies, Adjust notes, questions and proposals as they happen; folded to the result once it finishes), a box to Adjust, Stop, Answer or send a follow-up, cost so far, and Open. Starting a run from your own Regent chat adds one line there (Handed to …, its status and newest line) with Open in Team chat
• Hand-offs between teammates — a run can hand a piece of its work to another teammate; the transcript shows Handed "…" to Max linking to the new job, which runs as its own post marked from Ada. Anyone who can make changes on the business can answer a run's question, sign-in or secret card; a permission card stays with the person who asked (or an admin), and the card says so for everyone else
• Workspace & Daily Brief attention badges — open cards and questions increment the badge on Runs in the Regent workspace header, fold into the Proposed Changes badge in Books, send an agent notification, and appear on the Daily Brief tile
• Per-run cost & execution tracking — every run tracks prompt/completion tokens, tool calls, turn count (executions), engine/account label, and USD cost, rolled up into monthly role spend caps

How it works

When a job is triggered from Agent Team, Regent, a scheduled SOP workflow, or the API, DayZero queues an agent run on the durable Temporal agent worker. As the agent works, every thought, tool call, proposed change, and card is appended to the event log and streamed live to the drawer and chat card. If the agent hits a question or permission boundary, the run pauses in Needs you (needs input) and notifies you; once you answer (or approve/reject its proposed changes), the worker signals the run and it resumes from where it left off.

flowchart LR
  start["Queued -> Running"] --> step["Stream tool calls & reasoning"]
  step --> check{"Needs human input?"}
  check -->|"Card or question"| park["Needs you (paused)"]
  park -->|"Answer / Allow"| step
  check -->|"Gated write"| prop["File in Proposed Changes"]
  prop --> step
  check -->|"Done when met"| done["Done (Result & Close step recorded)"]

How to use it

1. Open Agent Runs in the Regent workspace, or click Open / Answer on a run card inside Regent or Notifications.
2. Use the filter pills (Open, Needs you, Done, Failed, All) and the Every role dropdown to locate a run. Rows show the run title or brief, role name, waiting card count, proposed change count, status pill, engine chip, start time, and cost.
3. Click any row to open the Run detail drawer on the right.
4. Expand Brief at the top of the drawer to review the instructions and Done when check given to the agent, or expand Thinking blocks in the transcript to see its reasoning.
5. If the run is in Needs you:
   • On a Permission card, review the tool name, reason, and command, then click Allow, Allow always (auto-approves matching requests for the rest of the run), or Deny.
   • On a Question card (or the yellow needs your answer form when a run ended its turn with NEEDS INPUT:), click one of the suggested option buttons or type your reply and click Answer.
   • On a Sign in card, click Open sign-in page, complete sign-in in the new tab, and click I signed in (or Skip).
   • On a Secret card, paste the one-time credential and click Send (it is handed to the run once via short-lived cache and never stored on the database row).
6. Click any Proposed link in the transcript (or under Proposed changes at the bottom of the drawer) to review and approve or reject the run's ledger edits in Proposed Changes.
7. To halt an active run, click Stop in the drawer header (or under the job in Team chat) and confirm Stop run.

Pro tips

• Check the Needs you filter first thing in the morning — answering a parked question or permission card resumes the run immediately so the agent can finish before you review Proposed Changes.
• Stopping a run is always safe: the agent stops after its current step, and any edits it already queued stay in Proposed Changes for you to approve or reject at your leisure.
• You can also manage runs directly from Regent in plain English: ask "how is the bookkeeper run going?", "tell the agent to use the September Chase statement", or "stop the bookkeeper run".
• If a run is interrupted by a server deploy or worker restart, DayZero's Temporal workflow automatically resumes the session (--resume) unless the turn already performed an outward action, preventing duplicate side effects.

In-depth guide

Run statuses and outcomes

Inline card types (AgentCard)

When a run needs help from a person, it opens one of four card types inline in the transcript:

Connected services in a run

Every server a business connects under Integrations → MCP is available to its agents unless an admin turns off Agents can use this server on that connection. In the transcript, a call to one of these servers is labelled with the service name and the tool, plus an External tag. Reading from a service runs straight away; anything that would change data in the service is filed in Proposed Changes first, and approving it runs exactly the call the agent asked for. A Read only role cannot make those calls at all. What a service returns is outside data: agents treat it as information, never as instructions.

Close checklist integration

When a run is launched for a job linked to a Monthly Close Checklist item, completing the run automatically records the close step (Done when successful, Blocked on failure) so the close checklist and the Regent workspace stay in sync.