Skip to content

Not a Chatbot

Jump to bottom
Dominic edited this page Jul 10, 2026 · 1 revision

This is not a chatbot

Warden is an autonomous agent system. The model you're talking to is the orchestrator; it has tools, sub-agents, and filesystem access. It responds best to actionable requests, not open-ended conversation.

Chatbot vs. agent

✗ Chatbot-style asks

  • "What do you think about Rust?"
  • "Tell me about microservices."
  • "Can you help me with my project?"
  • "Let's talk about our roadmap."
  • Outcome: vague, conversational, no action taken.

✓ Agent-style asks

  • "Read src/auth.ts and tell me if there's a timing-safe comparison missing."
  • "Screenshot example.com and tell me if the hero is above the fold."
  • "Convene The Council on this: should I use Postgres or SQLite for a single-user app with 50K rows?"
  • "Open my last 10 unread emails and summarize each in 1 line."
  • Outcome: tools run, files read, verdicts returned.

The five principles

  1. Be specific about the target. Include the file path, URL, or input data. "Fix the bug" → "Fix the bug in src/auth.ts:142 where the JWT expiry check is missing."
  2. One ask per turn is fine; parallelize independents. If two things don't depend on each other, say "X AND Y in parallel" — Warden runs them concurrently.
  3. Read BLOCKED messages. When a direct tool call fails, Warden says BLOCKED with the reason — read the error and redirect. Background Atlas jobs handle this themselves: the orchestrator catches failures, reworks the prompt, and retries before ever bothering you.
  4. Watch the verbose bar. Below the composer. Shows what Warden is doing right now. If it's calling tools you didn't expect, hit Stop.
  5. Use The Council for high-stakes, not for lookups. The Council is 3–10× slower and token-heavier. Reserve it for consequential decisions. Use direct tool calls for everything else.

The mental model

Treat Warden like a capable colleague you're briefing, not a search engine you're querying. Give it context, a target, and a definition of done:

Read the deploy log at logs/dockbox.log, find any error lines
from the last hour, and tell me which component is most likely
failing. If you can't tell from the log, say so — don't guess.

It'll read the file, grep for errors, analyze, and either answer or say it can't tell. That's the loop.

What it won't do well

Open-ended brainstorming without a target. Hypotheticals with no verification path. Anything that needs "human judgment" without grounding in something it can read or run. Long conversations that drift from the original ask — the context window is finite.

For history-heavy asks, Warden uses Mercury: a rolling summary plus RAG over older conversation turns, injected into every prompt. Configure it in Settings (Off / RAG / Summary / Full). It helps, but it's not perfect — be specific about time ranges or message IDs when precision matters.

If you want to brainstorm, point it at something concrete: "Read this PRD and tell me what's missing." That gives it a target.

Home

Clone this wiki locally