Agent-Shape Triage
Before spending on a task, estimate four properties of the work itself and let them pick the operating point: chat, one agent, a team of agents, or no AI at all. The vault's simplicity cluster already says default small and justify the climb (Agentic Simplicity) and that agent count is a design variable rather than a free dial (Bounded Fan-Out). Neither says how to decide on a specific task, in advance. This is that missing step: a pre-flight estimate cheap enough (the source says a minute) that running it costs less than guessing wrong.
The four questions, in the source's ordering:
- Size — is the task bigger than one agent can hold at full quality? Not "does it fit" but "does it fit without the context window filling and quality sagging." The source's calibration: a calendar "fits in a corner of a context window"; a quarter of email probably fits; a hundred documents "might scale out"; a thousand "definitely would."
- Independence — can the parts be done without knowing what the other parts did? A pile of documents splits cleanly because "one reader agent can read any given document and they never need to talk." Note the source's honest caveat: coding sometimes splits and sometimes doesn't, and which one it is depends on how you told the agent to organise the files — independence can be a property you create, not only one you find.
- Separation of concerns — do any parts need to be done by a different mind? A critic who didn't write the draft; an overview written by someone who didn't do the reading. If you're asking this, you are looking at a team (Team-Forming Constraints).
- Checkability — is checking an answer much cheaper than producing one? A test suite, an exit code, a source document you can point at. This is the question that gates whether extra spend converts into results at all (Repeated-Sampling Scaling).
The verdicts: small → chat. Fits one window, splits nothing, checks its own work → one agent with a goal. Exceeds one window, or has parts that need separate minds → a team. And the fourth, which the source calls the one that saves the most money — a judgment call you need to sit with → no AI at all (Cognitive Offload Cost).
The claim that makes this more than a checklist is why these four: they are properties of the work, not of the tooling. Models, harnesses, and team patterns all churn; "is this bigger than one mind can hold" and "can I check the answer cheaply" do not. The source's framing — "in a market where everything is disposable, this test is like a buy it for life purchase."
Claims
- The four estimates describe the work, not the tools, and therefore outlive the tools. principle — durable, and the load-bearing justification for the whole page. Size, independence, separation of concerns, and checkability are facts about a task; they were true of human teams before agents existed (see Team-Forming Constraints's audit and peer-review lineage) and will survive the current generation of harnesses. The source chose them on exactly this criterion.
- Estimate the shape of the work before choosing the machinery — the common failure is not bad tooling but pointing good tooling at the wrong shape. principle — durable: the mismatch is a category error, not a capability gap. The source's framing device is that 1.6M agents registered at OpenClaw's peak and most did nothing, because "everybody bought thinking, but no one knows what to point that thinking at." (The figure is the source's claim — see the distillate's Key claims.)
- Run the four questions as a fast triage rather than a rigorous analysis — the point is to get into the work, not to estimate. best practice — context: an operator with a real task and a bias toward either over- or under-engineering. The minute-long budget is the whole design: a more careful estimate would cost more than the wrong answer. "You're not sitting there estimating work for the sake of estimating. That's shaving the yak."
- Ship every verdict with a next step attached; a verdict without a way forward is just more homework. best practice — context: any tool or ritual that classifies work for a human who then has to act on it. Applies to firehose's own triage surface as much as to the source's tool.
- Do the minute of thinking yourself first, then check the tool — and treat a disagreement between your instinct and the tool as the learning signal. best practice — context: an operator deliberately building an instinct they don't yet have. Inverts the usual tool-first flow at a real cost in speed, and is worth it only while the instinct is still forming; the source's stated payoff is that agreement lets you "move with confidence" and disagreement is "probably something the tool is seeing that you may not be recognizing." See Cognitive Offload Cost — the reason to keep the human estimate in the loop is that the judgment is what atrophies.
- Independence can be manufactured, not just measured: whether a coding task splits depends on how you told the agent to organise the files. observation — the source's caveat, and the one place the four questions stop being pure observations about a fixed task. It means a "no" on independence is sometimes a design prompt rather than a verdict.
- The tool form takes two more inputs beyond the four work-shape estimates — how often the cost recurs, and what a good answer is worth. observation — the source's "money dials," held rather than developed here; they turn a shape classification into a spend decision. See the distillate's held themes.
Related
- Agentic Simplicity — the disposition this page turns into a procedure: "add complexity only when it demonstrably improves outcomes" tells you to justify the climb; agent-shape triage is how you decide, on a given task, whether the climb is justified. The fourth verdict (no AI) is that page's "sometimes the right answer is not to build an agentic system at all," made a first-class outcome.
- Team-Forming Constraints — where two of the four questions (size, separation of concerns) point: the only two limits that justify forming a team at all.
- Repeated-Sampling Scaling — what the checkability question is really asking. Checkability is not hygiene; it is the switch that decides whether more spend converts into more results.
- Bounded Fan-Out — the next decision down. Triage says whether a team; bounded fan-out says how many. Note the axes differ: attempts scale smoothly inside a verifier, agents do not.
- Workflows vs Agents — the sibling fork, decided on a different axis: that page routes on who holds control (predefined path vs model-directed); this one routes on the size/independence/ checkability shape of the work. A task can be agent-shaped here and workflow-shaped there.
- Incremental vs Discontinuous Tasks — the vault's other route-by-task-shape taxonomy, and a useful contrast: it predicts whether a model will succeed; this one predicts what cardinality to spend at. A discontinuous task and a "no AI" verdict often name the same work from two directions.
- Resolution-Typed Tasks — the same typing move one level down: that page types tickets inside an already-decided plan by how each resolves; this page types the whole task before there is a plan. Its HITL type and this page's "no AI" verdict are close cousins.
- Cognitive Offload Cost — why the fourth verdict is not a consolation prize: the source's report is that people "overindex on what they delegate to AI," and the reminder to not delegate may be the tool's most valuable output.
- Model-Tier Routing — the orthogonal spend lever: triage picks the shape, tier routing picks which model runs each role inside it.
- Total Cost of Inference — the bill the estimate is trying to bound, and a reminder that the cost it should be priced in is attention as much as dollars.
- Distillate: 1.6M agents registered for OpenClaw and did NOTHING.