Reference

Agents

Agents are your governed digital workforce. Each one runs a published, immutable version, and every side-effecting action it proposes waits for human approval.

The agents list

The Agents page lists every agent in your organisation. Status tabs across the top filter by lifecycle state (Draft, Active, Paused and Archived), and the default view shows Active agents, so archived ones stay out of the way until you ask for them.

• Filter by name (search) and by type, and sort by name, status, current version, last run or created date. Sorting and filtering happen on the server, and every control is reflected in the URL, so a filtered view is shareable and bookmarkable.
• Each row shows the agent's name and slug, its type, status, current published version, and the status and time of its last run.
• Viewing the list requires the agents.view permission; the New agent button only appears if you hold agents.create.

Creating an agent

Creation is a two-step page. First, choose a starting point: one of the eight library presets, or Start blank.

• Library presets pre-fill a working specialist configuration (objective, instructions and suggested tools) for a given type (Techshift Ops, Documentation, Research, Content, Reporting, Support, Sales or Compliance). A preset's suggested tools are highlighted in the version editor's Tools tab; you still assign them there before publishing.
• Start blank gives you an empty definition with the default policies. You pick the closest library type and write the objective and instructions yourself. Every agent maps to a library type even when it starts blank.
• Then name your agent. The slug is optional — a URL-safe identifier, unique within your organisation, generated from the name if you leave it empty. A description is shown in listings.

The AI Agent Builder

The Builder is a chat-first way to author an agent by describing what you need in plain language. It interviews you, drafts an objective and instructions, suggests tools (showing each tool's risk level and side-effect flag) and proposes an approval policy. All of it is presented as a live draft preview as the conversation progresses. If voice is configured, you can also speak your message with push-to-talk; the transcript is shown for you to confirm or edit before it is sent.

• The Builder never auto-publishes. The draft preview carries a hand-off link into the same version editor used everywhere else, where a human reviews and publishes through the unchanged approval gate.
• It can also draft workflows (ordered, multi-agent sequences), handing them off to the workflow editor for human review and activation.
• Builder sessions and turns are gated by agents.create, and assistant usage is metered and capped by your plan and per-session limits.

Trust score

Every agent carries an advisory trust score: a single figure, computed from the agent's audit, run and evaluation evidence, that helps you judge how much it has earned your confidence. A compact gauge appears on the agent overview, deep-linking to the full breakdown.

• Scores fall into three bands: good (75 and above), caution (50–74) and poor (below 50). The card also shows a confidence percentage and the trend since the last computation.
• The score is read-only and explainable: the full panel breaks it into components, each linked to the evidence that produced it. It is recomputed on a schedule, not on demand. Until the first nightly sweep runs for a fresh agent, the panel shows an empty state rather than a guessed figure.
• The trust panel sits behind agents.view — if you can see the agent, you can see its trust.

Trust is advisory. It informs your judgement; it never relaxes the approval gate or lets an agent act autonomously.

The agent overview

An agent's overview page is its hub. From here you can:

• Run the agent (requires runs.execute, and only while it is active with a published version) or Dry-run it in the sandbox. The agent reasons for real, but every side-effecting and integration tool is simulated, with no approvals, no real actions and no impact on your plan limits.
• Open the draft or start a new draft version (agents.update), and edit the agent's metadata.
• Browse the full version history, compare any two versions field-by-field, and review recent runs (runs.view).
• Jump to Performance, Evaluations, Trust and Earned autonomy, or submit the agent as a marketplace template (agents.publish) — a config-only snapshot, never credentials or tenant data.

Published versions are immutable: every change becomes a new version, and the whole lifecycle is recorded in the audit log.

Next steps