-
Notifications
You must be signed in to change notification settings - Fork 3
The Agents
Warden is one orchestrator plus a handful of specialist sub-agents. The orchestrator is the model you're talking to, but it is explicitly a router and prompt engineer — it does not write code, edit files, or do research itself. Instead it understands what you want and delegates to the right sub-agent. Each sub-agent runs in its own context window and reports back as a chat message.
You → Orchestrator (plans + delegates) → Atlas / Byte / Dexter / Iris / Artemis / The Council
Multiple sub-agent calls in the same turn run in parallel. The verbose bar below the composer shows which one is live.
| Agent | Specialty | Example |
|---|---|---|
| Atlas | Browser + research: opens URLs, takes screenshots, fetches HTML, clicks buttons, reads pages. Multiple Atlas jobs in one turn run in parallel. | Open example.com and take a screenshot. |
| Byte | Compute: runs python3 / bash / node one-liners and hands back stdout. | Compute the factorial of 10 using python3. |
| Dexter | Time + scheduling: date/time questions, recurring or one-shot scheduled tasks. | Remind me at 3pm to check the deploy status. |
| Iris | Email: lists your inbox, counts unread, searches, and (if enabled) drafts and sends. | List my last 10 unread emails and summarize each. |
| Artemis | Read-only auditor: reads files, greps, reviews code or conversation. Also seats The Council. | Review src/auth.ts for any security issues. |
| The Council | 3 Artemis seats (Skeptic, Pragmatist, Synthesist) deliberate in parallel until they agree or hit max rounds. 1–3 min, token-heavy. | Convene The Council on this: should I use a monorepo for a 2-person team? |
The orchestrator is a prompt engineer and delegator. You describe the task; it picks the right specialist, writes a precise instruction (including file paths, URLs, and expected output), and emits the tool call. The matching sub-agent runs and posts its result back.
Delegation is supervised: the orchestrator checks in on running Atlas jobs every 45 seconds (elapsed time, tool calls, last action), stops any that are stuck or looping, and when a job fails it reads the output and automatically retries with a reworked instruction — you only hear about failures it can't recover. The verbose bar at the bottom of the chat shows which agent is live and which tools are running right now.