abc-agent-management

name	abc-agent-management
description	Functional-analysis protocol for iteratively improving subagent performance. Adapted from Antecedent-Behavior-Consequence (ABC) analysis in applied behavior analysis. Use when delegating to a subagent, when an agent returns a poor result, when a pattern of similar failures appears across runs, or when coaching a prompt toward reliability.

ABC Agent Management

Treat subagent failures as behaviors with antecedents (the prompt/context that produced them) and consequences (the feedback loop that sustains them), not as traits of the agent. Fix the environment, not the "agent."

Mapping

ABA construct	Agent equivalent
Antecedent (immediate)	The brief given to the agent, constraints, exemplars
Setting event (distal)	System prompt, prior turns, tool set, context-window state
Behavior (observable)	Tool calls, output, artifacts, the scope the agent chose
Consequence	Your response: accept, correct, retry, reject, downstream effect
Function	What "need" the behavior met (escape, completeness theater, pattern-match, reinforcement history)
Reinforcement	Anything that makes the behavior more likely next time
Extinction	Removing what maintained the behavior; expect a burst before improvement

Protocol

1. Operationally define the behavior. No trait words ("sloppy", "lazy"). State what was observed, measurable. "Emitted 200-line plan, 0 tool calls, on an implementation task" — not "didn't do the work."

2. Record A, B, C before intervening. Skipping this is how you fix the wrong thing.