Should your agent workflow use a protocol at all?

说明与用途

在代理工作流中使用协议的决策应基于 Git 推断的风险信号，而非代理自报。Agateon 计算仪式深度，并在声明的级别低于计算值时阻止任务。

问题

Deciding whether to add an orchestration layer to an agent workflow can be costly if done incorrectly – too much process wastes agent time, too little leads to incomplete work.

结果

In a sample of 37 tasks, the protocol added a median of 21 hours of calendar time but only 3 real gate blocks, demonstrating that gates rarely stop work but provide coverage of hidden failures.

工作流

• 1. Stage changes for the task
• 2. Run agate-risk-score.py to compute tier and evidence lines
• 3. Declare desired ceremony depth (thin/standard/full) in task metadata
• 4. Gate checks declaration against computed tier – blocks if too low
• 5. Execute task phases, generating artifacts and logs
• 6. Review gate outcomes and retry if necessary

证据

The risk score exists because ceremony depth used to be self‑reported, and self‑reported complexity drifts toward whatever the agent thinks you want to hear.

来源

Across 37 tasks in agate-workspace/tasks/: median wall‑clock span ~21 hours, 13 commits, ~6,700 phase artifact lines.

来源

Of 119 gate runs recorded, 115 ended in the pass code for that phase, 3 exited 1 – real blocks.

来源