Skip to content

Latest commit

 

History

History
29 lines (19 loc) · 2.3 KB

File metadata and controls

29 lines (19 loc) · 2.3 KB

Standard HOOTL/HITLFE/HIC task classification for AI agents

Status

Accepted

Context

AI agents operating in the orchestrator must decide when to proceed autonomously, when to pause for human approval, and when to surface a task and wait entirely. Without a shared classification system, agents either over-interrupt (training the human to click through without reading) or under-interrupt (missing high-stakes decisions mid-flight).

Decision

Three-tier classification:

  • [HOOTL] — AI executes autonomously; human audits after-the-fact. High confidence, deterministic outcome. Agent executes without prompting.
  • [HITLFE] — AI handles routine prep, escalates exception to human for approval. Agent completes all prep work first, then pauses with a structured context summary (what was done, why input is needed, potential impact) before proceeding.
  • [HIC] — Human-in-Command. Task requires creativity, judgment, or empathy that AI cannot substitute. Agent surfaces the task and waits.

Agents follow Prepare-Then-Prompt: no asking permission before acting on [HOOTL] tasks. Multiple [HITLFE] and [HIC] items are batched into a single daily briefing rather than scattered interrupts.

Considered Options

  • Interrupt-on-uncertainty — ask the human whenever confidence drops below a threshold, without doing prep work first. Rejected: creates constant low-signal interrupts and shifts cognitive load onto the human before they have enough context to decide.
  • Fully autonomous with human review after — agent completes everything, human reviews in post. Rejected: too late to catch high-stakes or hard-to-reverse actions mid-flight; removes meaningful human control over consequential steps.
  • Explicit per-task approval prompts before every action — rejected: high friction, trains users to click through without reading, and collapses [HOOTL] and [HITLFE] into one noisy mode.

Consequences

  • Agents must tag each task node at planning time; ambiguous tasks default to [HITLFE] rather than [HOOTL].
  • The daily briefing pattern requires agents to track which tasks pended human input across a session and surface them together at natural stopping points.
  • This classification is an orchestrator-level convention; project submodules may adopt it internally but are not required to.