Skip to content

Configuring an agent

The agent's detail page has tabs across the top: Dashboard, Configuration, Skills, Files, Runs. This page covers the Configuration tab — what each card does and when to touch it.

The page header (above the tabs) is where the live controls live: Run Heartbeat, Pause, Resume, and Delete.

Properties card

In the Configuration tab, the Properties card holds the basics:

Field Effect
Name The display name. (Also editable from the header.)
Role Switches the role label. It does not rewrite the system prompt — edit that yourself if you want it changed.
Reports to The agent's manager in the org chart. Read by the escalation playbook.
Adapter type default, llm, claude_local, codex_local, docker, lambda, or remote. default follows the company's configured default adapter (resolved each run) — pick it for portable agents. Switching it changes which configuration cards appear below.

Click Save Properties when you change anything here. There are also small cards for Home Directory (where this agent's skills and files live on disk) and Direct Reports.

AI Provider card (all adapters)

Every agent has an AI Provider card, regardless of adapter:

  • Provider — pick from the company's connected providers, or leave it None to fall back to the adapter-specific config below.
  • Model override — the specific model to use. It autocompletes from the provider's discovered model list; leave it blank to use the provider's default.
  • Search filters (Perplexity providers only) — Search Recency Filter (last hour/day/week/month) and Search Domain Filter (a comma-separated allow/deny list of domains).

Click Save Provider.

For an llm agent, this is usually all you need: pick a provider, save, and set the heartbeat (below).

Heartbeat (every adapter)

Each adapter's configuration card includes the heartbeat controls:

  • Enable automatic heartbeats — the checkbox that turns the agent on. Unchecked means it only runs when you click Run Heartbeat.
  • Interval (seconds) — how often it wakes up. Minimum 10; 60 is a sensible default. See the heartbeat page for tuning.

Click Save Configuration to apply.

The llm adapter card

For agents that call an LLM provider directly. Beyond the heartbeat controls, this card carries the fallback model config used when no Provider is selected above:

  • Provider — OpenAI, Anthropic, Perplexity, or Custom (OpenAI-compatible).
  • Model, API Key, Base URL — the connection details.
  • Max Tokens, Temperature — generation tuning.
  • Prompt Template — the agent's system prompt. This is the single source of truth for the prompt, used by every adapter type. Edit it to:
    • Tighten or widen scope ("you only handle frontend issues").
    • Steer escalation ("anything touching payments needs an approval").
    • Encode conventions ("we use /blocked-by, not /relates").

Prefer connecting a real provider in the AI Provider card over filling in keys here — it's reusable across agents and testable with Test Connection on the Providers page.

The claude_local and codex_local cards

For agents that run a coding CLI on the server. Instead of provider keys you get a Session card and a Configuration card:

  • Working Directory — the path on the server where the CLI works.
  • Model — the model the CLI should use.
  • Max Turns Per Run, Timeout — run limits.
  • Skip permission prompts — a dev-only toggle.

Sessions persist between heartbeats by default, so the agent continues the same CLI conversation across runs. Test Environment runs a quick check against the local CLI; Clear Session (shown when a session exists) starts the next run fresh.

These adapters are operator-managed — they need the CLI installed and authenticated on the server. See picking adapters.

The docker and lambda cards

For containerised or HTTP-callable agents. Docker takes an image, pull policy, and mount points; Lambda / HTTP takes an endpoint URL and auth header. Both are operator-level concerns — most board users won't open these. See picking adapters.

Skills tab

Two columns: skills the agent can see, and skills available to add. Add by clicking; remove the same way. Skills come in two flavours:

  • Filesystem skills — richer SKILL.md + skill.yaml pairs on disk. Tell an agent it has the hire-agent skill and it reads that content as extra context.
  • DB skills — short structured entries in the company catalog.

See writing a skill to author your own.

Files tab

Lists files in the agent's home files/ directory. Upload, download, and delete here — handy for handing the agent reference material.

Runs tab

A reverse-chronological list of every heartbeat run for this agent: model, tokens in/out, cost, status, start/finish times, and the linked issue if any. Click a run to open its full transcript and event stream.

Proposed prompt revisions

If the optional self-improvement worker is enabled, you may see a Proposed prompt revisions section near the bottom of the page. When the worker spots a pattern in the agent's recent runs, it drafts a new system prompt and offers it here. You can Accept this proposal (mints a new prompt version and makes it active), view a side-by-side diff first, or Reject it with a reason. The page also keeps a version history of the prompt with Save and promote and Restore controls, so a prompt change is always reversible.