Decision: Agent Dottie (Documentation Subagent)
Decision: Agent Dottie (Documentation Subagent)
Decision record for creating a fourth agent in the operating model. Made 2026-03-07 by Ava and Lucas.
Context
At 133K tokens into a session, documentation operations consume 5,000-15,000 tokens each. Creating a note requires reading the style guide, reading related notes, composing HTML, calling create_note, and verifying the result. Updating a note requires reading current content, modifying it, calling update_note. Each operation expands the main session conversation, consuming the most valuable resource: context window space in Ava's session.
The main session is where strategic decisions happen, sprint planning occurs, and agent coordination takes place. Every token spent on mechanical doc operations is a token not available for higher-value work.
Decision
| Aspect | Choice |
|---|---|
| Action | Create a documentation agent ("Dottie") to execute doc tasks in a separate context |
| Name | Dottie (from "docs") |
| Role | Librarian -- executes doc updates, content audits, quality tracking |
| Access | pal-e-docs read/write + Forgejo read-only. No code, no repo writes. |
| Spawned as | General-purpose subagent with Dottie personality injected in prompt |
| Boss | Ava. Dottie never makes strategic decisions. |
Token Economics
| Operation | Inline (before Dottie) | With Dottie | Savings |
|---|---|---|---|
| Create a note | ~5,000-8,000 tokens (read style guide + related notes + compose + create + verify) | ~300 tokens (spawn prompt + summary response) | ~95% |
| Update a note | ~3,000-6,000 tokens (read current + modify + update) | ~300 tokens | ~90% |
| Content audit | ~10,000-15,000 tokens (read many notes, compile findings) | ~300 tokens | ~97% |
| Batch doc updates (3+ notes) | ~15,000-25,000 tokens | ~300 tokens | ~98% |
Key insight: The savings come from context isolation. Dottie's full conversation (reading notes, composing HTML, making MCP calls) happens in a separate context window. Ava only sees the spawn prompt and the summary.
Why Not Earlier?
- Subagent spawning was not available until recently
- Earlier sessions were shorter and doc operations were fewer
- The three-agent model (Ava, Dev, QA) was sufficient when docs were simpler
- As the corpus grew to 256 notes and doc operations became more frequent, the cost became visible
Constraints
- Dottie never makes strategic decisions -- presents options to Ava
- Dottie never writes code in repos
- Dottie never creates or closes Forgejo issues (that is Ava's job)
- Dottie always follows note templates and html-style-guide conventions
- Dottie always reports what was created/updated/changed
Impact on Operating Model
The three-agent model becomes four agents. Rule 4 ("Ava owns docs") is refined: Ava directs doc operations; Dottie executes them. The separation preserves the principle that agents don't make strategic decisions about documentation -- Dottie is purely mechanical execution.
Related
agent-dottie-- Dottie's personality definitionagent-workflow-- the operating model (updated to include Dottie)agent-ava-- Dottie's boss