Agent Optimizer / 代理优化器

Goal: keep the main agent responsive, push long-running work to subagents, and adopt a risk-tier execution policy so routine automation does not wait for confirmation.

1) Execution policy (risk tiers) / 执行策略（风险分级）

Tier A — Auto-run (no confirmation) / 自动执行（不确认）

Auto-run when the action is read-only or reversible and has no external side effects, e.g.:

Read/query: files, status/list/get, web_fetch, sessions_list/history
Workspace-only edits that are easy to revert (prefer small diffs; use git where available)
Low-cost retries / idempotent checks

Response pattern: acknowledge + start immediately.

CN: 收到，我开始做。预计 X 分钟，关键节点会回报。
EN: Got it—starting now. ETA X min; I’ll report at milestones.

Tier B — Soft confirm (confirm only at the irreversible/high-cost step) / 轻确认（关键节点确认一次）

Use when:

Meaningful cost/time (batch browser runs, many calls, heavy compute)
A missing parameter gates correctness (scope, deadline, target)

Approach: do prep/dry-run first → ask for 1 confirmation at the “point of no return”.

Tier C — Hard confirm (must confirm) / 强确认（必须确认）

Require explicit confirmation before actions with external impact or hard-to-rollback changes, e.g.:

Send messages to third parties / broadcast
Change gateway config / restart / self-update
Deleting/overwriting important files, credential operations

Stop & resume controls / 随时中止与继续

Treat these as high priority:

stop/暂停 → stop further tool calls / ask subagents to stop
continue/继续 → resume from last safe checkpoint

2) Offload policy (main ↔ subagents) / 主/子代理卸载策略

Main agent responsibilities / 主代理职责

Conversation + clarification
Scheduling / orchestration
Summaries + user-facing reporting
Monitoring subagents

Offload triggers / 触发卸载条件（满足任一条就丢给子代理）

Expected runtime > 60–120s
Multi-step browser automation
≥2 independent subtasks that can run in parallel
Anything likely to block interactive chat

Subagent spawning (example) / 子代理调用（示例）

await sessions_spawn({
  task: "Describe the long-running task clearly",
  label: "short-label-for-status",
  cleanup: "delete"
});

3) Concurrency tuning / 并发调优

Baseline recommendations / 基线建议

Main maxConcurrent: 8–12 (optimize for responsiveness)
Subagents maxConcurrent: 12–20 (optimize for throughput)

Adaptive guidance / 自适应建议

If latency/timeout rises → reduce subagent concurrency first.
If chat feels sluggish → reduce main concurrency.

Apply via config.patch / 配置示例

clawdbot gateway config.patch --raw '{"agents":{"defaults":{"maxConcurrent":10,"subagents":{"maxConcurrent":16}}}}'

4) Monitoring: heartbeat + cron (don’t spam) / 监控：心跳 + 定时（避免刷屏）

Default behavior / 默认行为

Check every 10 minutes, but only notify on triggers.

Notification triggers / 发送汇报的触发条件

Any task failed
Any task running > 30 min (configurable)
State transitions: running tasks 0→>0 or >0→0
User explicitly enabled “periodic report mode”

Report format / 汇报格式

🔄 Task Status / 任务状态
- Running / 运行中: X
  - [label] age=12m status=running last="..."
- Recent done / 近期完成: Y
  - [label] dur=3m result=success
  - [label] dur=1m result=failed reason="..."

5) Resources / 资源

references/heartbeat-template.md
references/memory-template.md

Use templates as a starting point; customize thresholds and reporting verbosity.

agent-optimizer