Agent-Computer Interface

ELI5 - The Vibe Check

This is the difference between handing someone a giant printed manual and handing them a clearly labeled control panel. Same underlying machine, same capabilities, but one is designed for a human to fumble through page by page and the other is designed for an agent to read fast, parse reliably, and act on without guessing.

Real Talk

Agent-computer interface (ACI) is the design discipline of shaping tools, file formats, and command outputs specifically for consumption by an LLM agent rather than a human, popularized by SWE-agent research. A good ACI returns concise, structured, parseable feedback, such as line numbers, diffs, or error summaries, instead of raw human-oriented terminal output, which measurably improves an agent's task success rate.

Show Me The Code

# Human-oriented output (hard for an agent to parse reliably)
$ grep -n "TODO" app.py
14:    # TODO fix this later

# Agent-Computer Interface output (structured, unambiguous)
$ search_code --pattern "TODO" --file app.py --format json
{"matches": [{"line": 14, "text": "# TODO fix this later"}]}

When You'll Hear This

"We redesigned the tool output for the agent, not for humans, and the success rate jumped." / "Half of ACI work is just returning line numbers instead of raw text."