Resolution-Typed Tasks
When you decompose a plan into tickets, tag each one not just with what it is but with how the open item resolves and whether a human must be present while it does. The source's Wayfinder skill labels every ticket on two axes:
- Presence axis — HITL vs AFK. A HITL (human-in-the-loop) ticket is worked with a human who speaks for themselves; it "only resolves through that live exchange, and the agent never stands in for the human's side of it." An AFK (away-from-keyboard) ticket is driven by the agent alone — hand it off and come back. The load-bearing rule: an agent that answers its own HITL questions (grilling itself) has broken the type.
- Mode axis — how it resolves. Research (AFK): read primary sources / docs / local knowledge, write a markdown summary as a linked asset. Prototype (HITL): build a cheap artifact to react to when "how should it look / behave" is the question (Fidelity-Raising Prototype). Grilling (HITL): a one-question-at-a-time conversation — the default case. Task (HITL or AFK): manual work (config, provisioning, moving data into shape) that must happen before a decision can be made, where there's nothing to decide, prototype, or research but the plan is blocked until it's done.
The type is not decoration — it routes the work: it tells the orchestrator which skill or agent resolves the ticket, whether it can run unattended, and (via HITL) where a human gate is mandatory. It is the typing layer on top of an Agent Task Graph: the graph says B waits on A; the type says A is resolved by a background research agent or A needs you in the room.
Claims
- Type each open ticket by its resolution mode (research / prototype / grilling / task), because the mode determines which skill or agent resolves it and whether it can run unattended. best practice — context: decomposing a plan whose items resolve in genuinely different ways; the four-mode taxonomy is the source's, and its value is routing, so it pays off only when a plan actually mixes modes. A uniform to-do list doesn't need it.
- Mark whether a ticket is HITL (a human must be present, speaking for themselves) or AFK (agent alone) — and forbid the agent from standing in for the human's side of a HITL ticket. principle — durable: the whole point of a human-in-the-loop step is the human's own judgment; an agent simulating both sides (an agent grilling itself) defeats it, independent of tool. This is the failure mode that motivated the source's grilling fix.
- Grilling is the default resolution mode; research, prototype, and task are the exceptions you reach for when the unknown isn't "talk it through with the human." observation — the source's ordering: most planning unknowns resolve by conversation, so the taxonomy's other modes are named precisely so they don't collapse back into "just grill."
- A "task" ticket captures work that blocks a decision but contains no decision — provisioning, config, data shaping — the boring prerequisites AI can't resolve for you. observation — naming this type keeps it from being mistaken for a research or grilling node and stalling the map.
Related
- Agent Task Graph — the dependency substrate this types: blocks/blockedBy edges plus ownership; resolution-typed tasks add how each node resolves on top of what waits on what.
- Map-First Planning — the planning pattern whose map is built out of these typed tickets.
- Fidelity-Raising Prototype — the "prototype" resolution mode as its own primitive.
- Agent Supervision — HITL is a mandatory human gate; the type marks exactly where supervision is non-optional versus where the agent may run AFK.
- Leading Words — "HITL," "AFK," and the mode names are themselves compact terms the agent has priors for; typing a ticket is partly a leading-word steer on how to resolve it.
- Distillate: Skills v1.1: Wayfinder, the SDLC flow, and naming the artifact right