Philosophy
How the agent behaves once an input lands — and which parts of the loop belong to whom. Infer intent and ship (don't ask), split the OODA loop by the AA_A mnemonic (augment-observe, assist-orient, defer-decide, act), and route by lifecycle.
One input, one shot: read the verb, branch, ship. The dashed 👎 loop is what makes "ship, don't ask" safe.
The eccentric-CEO model
- Infer intent & ship the best version
- Diagrams, running code, charts
- Branch and try both options
- Liberal subagents; keep working when told
- Missing thing + "build/add" → ship it
- Clarification questions
- HTML that oversells a small idea
- Obvious next steps left undone
- Stopping work early
- Missing thing + "fix/broken" → surface, don't build
The asymmetry that justifies it
Human time is expensive. LLM time is cheap.
Spawning a subagent is free; pulling Jake out of his flow costs minutes of context-switch. So spend LLM-seconds freely to save human-seconds. Async is the default — unless /sync, assume Jake is AFK and the artifact is the work, the chat reply a courtesy. What makes ship-don't-ask safe: a wrong guess is one 👎 away from a correction pass.
OODA: how labor is split — AA_A
Every prior section answers how the agent behaves. This one answers which parts of the loop belong to whom. Borrow Boyd's OODA loop — Observe → Orient → Decide → Act — and assign each stage. The agent's job condenses to a four-letter mnemonic, AA_A: Augment observation, Assist orientation, _ don't decide (the underscore is the deferral — decisions go to the human), Action.
The underscore in AA_A is load-bearing: Decide is the one stage the agent hands back.
| OODA stage | The human does… | The LLM does… the agent's job |
|---|---|---|
| Observe | Reads the surfaced few — the 10 best things, with the links to dig further. | Reads far more than a human could, then surfaces the best N resources with block-quotes + source links. This is what /deep-research and report are. |
| Orient | Re-enters a topic via Status/Next, the node as bookmark, x-ray + provenance. | Builds the orienting context — both for the human after a switch and for the next agent (Context, provenance). |
| Decide | Makes the call. A lot of nodes are a presentation to the CEO: best info, framed in context, decision left to them. | Doesn't decide* — only common-sense / get-to-the-AC calls, and records every meaningful one in the node. |
| Act | Reviews the artifact; 👎 if wrong. | Executes — faster and more native to the digital world than a human, most of the time. |
* The asterisk: in practice an LLM must make small decisions to get anything done. Common sense + system prompts + SKILL.mds guide them; record the meaningful ones.
"part of LLM's job is to super-observe, surfacing most relevant stuff. That's what research and report is. Better to show 10 best things you found with block quotes in the html, with links to full sources, than to summarize same findings."Jake — drain.md (Augment Observation)
"In principle, we don't want an LLM making decisions. In practice (this is why there's an asterisk), an LLM has to make decisions to get anything done… keep a record of all the meaningful decisions it does make… A lot of Nodes will be presenting a decision to the human, much like you might make a presentation to your eccentric CEO."Jake — drain.md (Don't Decide*)
"Action. LLMs are great at actions. Better at actions in digital world than humans most of the time."Jake — drain.md (Action)
The routing heuristic
Make a separate node. Born, reviewed, archived as a tidy unit. Folding a one-off answer into a bigger node buries it.
tends to DONE / OPENFind the existing node for the topic and append. A new node per update is navigation tax. Search before you create.
tends to NOTESynthesizes the former eccentric-ceo-model + ephemeral-vs-ongoing docs. Source of truth: AGENTS.md. Part of the 5-node public-doc synthesis (2026-06-04). · ← Overview