Agent Gallery

The Agent Gallery gives teams curated starting points for common Dawn agents, with suggested prompts, model choices, skills, and source expectations.

The Agent Gallery is a set of reusable agent templates. Use it when you want a sensible starting point for a common role, such as support triage, PR review, release notes, product planning, customer research, or workspace operations.

Gallery agents are starting points, not locked presets. After creating one, review the prompt, model, skills, memory access, channels, and connected sources before inviting people to rely on it.

  1. Open Agents in Dawn Web.
  2. Choose New agent.
  3. Select a gallery template.
  4. Review the proposed name, provider, model, prompt, and defaults.
  5. Assign the connected sources the agent should use.
  6. Choose any memory access and skill assignments.
  7. Save the agent and test it in a private thread.

Start with one or two agents. A workspace with many overlapping profiles is harder to operate than a workspace with a smaller set of well-scoped agents.

A template can suggest:

  • agent name and description;
  • provider and model family;
  • system prompt shape;
  • default skills;
  • suggested memory behavior;
  • common source families;
  • whether the agent is best suited to chat, background work, workflows, or channels.

The workspace still controls what the agent can actually access. A template that expects repository context will not read repositories until a matching source is connected and assigned.

After creating a gallery agent, adjust it for your workspace:

  • rewrite the prompt in your team’s language;
  • narrow the source assignments;
  • add or remove skills;
  • choose whether workspace memory should be available;
  • pick a model that matches the job and cost profile;
  • bind the agent only to the channels or clients where it should respond.

Use the gallery to avoid blank-page setup, then make the agent specific enough to be reliable.

Validation Checklist

  • The agent answers from the intended sources and names them when relevant.
  • The agent does not answer from unrelated sources.
  • The prompt describes the role, boundaries, and output format clearly.
  • Memory access is explicit: none, all workspace memory, tag-scoped workspace memory, or personal memory only when acting for that user.
  • The model is allowed by the workspace AI provider policy.
  • Any channel binding or workflow usage has been tested before broad rollout.