ACI Design Principles

Keep agent actions simple and efficient, feedback informative but concise, and add guardrails that catch common errors before they compound.

Four principles for shaping an agent-computer interface around a fixed LM (weights unchanged; actions, docs, and feedback adapt):

Simple actions — few options, concise docs; avoid bash-style option explosions that demand demos or finetuning.
Compact, efficient actions — important operations (navigate, edit) should finish in as few turns as possible; do not force composing many micro-actions for one higher-order intent.
Informative but concise feedback — show substantive state/effects (for example the revised file window after edit) without dumping irrelevant detail that burns tokens under lost in the middle. On the failure path this means failures returned as actionable feedback: errors, timeouts and hangs reported in a form the model can act on.
Guardrails — catch common mistakes early (syntax checks, result caps) so recovery is fast rather than letting compound mistake amplification run.

Concrete SWE-agent instantiations: stateful file viewer, guardrailed edit tool, bounded search observations, collapsed observations, and action format enforcement. For browser agents, see lm-oriented web observations and token-efficient quote observations. Iterate by inspecting failed trajectories on a development set, then grid-search ACI knobs — the same spirit as failure-derived prompt tips. Aim toward automated ACI iteration so those inspections become a closed loop rather than only craft. Quantify each change with ACI component ablation: human-UI patterns (iterative search, full-file views, silent whole-file overwrites) often regress resolve rate even when they feel natural.

Sources