File size: 19,441 Bytes
2aeca9f aa0b221 2aeca9f aa0b221 2aeca9f aa0b221 2aeca9f aa0b221 2aeca9f aa0b221 2aeca9f aa0b221 2aeca9f aa0b221 2aeca9f aa0b221 2aeca9f aa0b221 2aeca9f aa0b221 2aeca9f aa0b221 2aeca9f aa0b221 2aeca9f aa0b221 2aeca9f aa0b221 2aeca9f aa0b221 2aeca9f aa0b221 2aeca9f aa0b221 2aeca9f aa0b221 2aeca9f aa0b221 2aeca9f aa0b221 2aeca9f aa0b221 2aeca9f aa0b221 2aeca9f aa0b221 2aeca9f aa0b221 2aeca9f aa0b221 2aeca9f aa0b221 2aeca9f aa0b221 2aeca9f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 | ---
title: "Planner-Executor: A Two-Agent Pattern for Reliable Email-Driven Automation"
emoji: π€
colorFrom: purple
colorTo: blue
sdk: static
pinned: true
tags:
- agents
- architecture
- multi-agent
- planner-executor
- email-automation
- claude
- agentic-ai
---
# Planner-Executor: A Two-Agent Pattern for Reliable Email-Driven Automation
> This pattern is implemented in **ClonAgent** ([clonagent.utopiaia.com](https://clonagent.utopiaia.com)) and has been running in production since July 2025.
> Source: [github.com/KikoCisBot/clonagent](https://github.com/KikoCisBot/clonagent)
---
## The Problem with Single-Agent Email Loops
The naive approach to email-driven AI automation looks like this:
```
email arrives β launch agent β agent reads, thinks, acts, replies β done
```
This breaks down quickly in production:
- **Long tasks block the inbox**: while Claude is deploying code (which can take minutes), new emails pile up unread
- **Context collapse**: a single agent session mixes "what should I do?" with "how do I do it?" β two very different cognitive modes
- **No retry logic**: if the action fails mid-way, the whole session is lost
- **No priority**: email #3 might be urgent but gets queued behind email #1
The Planner-Executor pattern solves all of this.
---
## The Pattern
```
βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
β Email Arrives β
ββββββββββββββββββββββββββββ¬βββββββββββββββββββββββββββββββββββ
β (poller tick, every 60s)
βΌ
βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
β SCHEDULER (Planner) β
β β
β Role: understand + plan β
β Input: email thread β
β Output: agent-tasks.json (list of structured tasks) β
β Duration: seconds β
β Blocks: nothing β
ββββββββββββββββββββββββββββ¬βββββββββββββββββββββββββββββββββββ
β writes agent-tasks.json
β (async β Scheduler exits)
βΌ
[2 min executor tick]
β
βΌ
βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
β EXECUTOR β
β β
β Role: execute pending tasks, one by one β
β Input: agent-tasks.json β
β Output: task results + email replies β
β Duration: seconds to minutes β
β Blocks: only itself (never the Scheduler) β
βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
```
Two separate Claude CLI sessions. Two separate prompts. One shared state file.
---
## Implementation
### The State File: `agent-tasks.json`
The only communication channel between Planner and Executor is a JSON file that lives in the agent's workspace:
```json
{
"agentId": "btc-signals",
"plan": "User requested BTC price analysis for the week. One task: generate report and reply.",
"tasks": [
{
"id": "a1b2c3d4-...",
"title": "Generate weekly BTC analysis",
"description": "Fetch price data, compute indicators, write summary, reply to thread",
"priority": 1,
"status": "pending",
"threadId": "18f3a2b...",
"from": "kikocisneros@gmail.com",
"createdAt": "2026-05-31T09:12:00.000Z",
"retries": 0,
"log": [],
"dependsOn": []
}
],
"updatedAt": "2026-05-31T09:12:05.123Z"
}
```
**Task statuses**: `pending` β `in-progress` β `done` | `failed` | `skipped`
**Priority**: `1` = urgent, `2` = normal, `3` = low
**`retries`**: auto-incremented on failure and recovery
**`log[]`**: timestamped progress entries written by the Executor during execution
**`dependsOn[]`**: IDs of tasks that must be `done` before this one is eligible
This file is readable by humans, inspectable at any time, and survives process crashes.
---
### The Planner Prompt
The Scheduler receives a tightly scoped prompt that constrains it to planning only:
```
ROL: PLANIFICADOR
Nueva comunicaciΓ³n recibida:
De: kikocisneros@gmail.com
Asunto: Weekly BTC report please
ThreadId: 18f3a2b...
Tu tarea (sΓ© rΓ‘pido y concreto):
1. Lee el email: python3 mail_client.py get-thread "18f3a2b..."
2. Lee el fichero de tareas actual: cat agent-tasks.json
3. Decide quΓ© tareas hay que hacer. Por cada tarea, aΓ±Γ‘dela con esta forma:
{
"id": "<uuid4>", "title": "...", "description": "...",
"priority": 1, "status": "pending",
"threadId": "...", "from": "...", "createdAt": "<iso>",
"retries": 0, "log": [],
"dependsOn": [] β IDs of other tasks that must be done first
}
- priority: 1=urgent, 2=normal, 3=low
- dependsOn: list task IDs in this file that must be "done" before this one runs
- Do NOT add tasks if threadId already exists with any status other than "failed"
- If threadId exists with status "failed", add nothing β the Executor will retry automatically
4. Actualiza el campo "plan" con tu anΓ‘lisis breve
5. Escribe el fichero actualizado
6. Sal cuando hayas terminado de planificar.
```
Key constraints on the Planner:
- **It never executes**. It only reads and writes `agent-tasks.json`
- **It deduplicates**: checks existing tasks by `threadId` before adding new ones
- **It models dependencies**: can express "deploy after tests pass" with `dependsOn`
- **It exits immediately** after writing the plan
---
### The Executor Prompt
The Executor picks up `agent-tasks.json` and works through it, respecting dependencies and logging progress:
```
ROL: EJECUTOR
Ejecuta las tareas pendientes de agent-tasks.json.
Proceso (repite hasta que no haya mΓ‘s tareas ejecutables):
1. Lee agent-tasks.json
2. Toma la tarea con status="pending" de mayor prioridad (1=urgente)
cuyo dependsOn[] estΓ© vacΓo o todas sus deps estΓ©n "done".
Si ninguna cumple esto, sal.
3. Actualiza status a "in-progress"; aΓ±ade al log[]: { "ts": "<iso>", "msg": "Iniciando: <tΓtulo>" }
4. Ejecuta la tarea; aΓ±ade entradas al log[] con progreso real durante la ejecuciΓ³n
5. Al terminar:
- status: "done" o "failed"
- result: descripciΓ³n breve del resultado
- updatedAt: <iso>
- Si retries >= 3: status="failed", result="Max retries reached: <reason>"
6. Si quedan tareas pending ejecutables, vuelve al paso 1
7. Cuando no haya ninguna tarea pending ejecutable, sal
Reglas:
- Marca "failed" (no crashees) si algo falla, continΓΊa con la siguiente
- El campo log[] es visible al usuario β ΓΊsalo para reflejar progreso real
- Usa mail_client.py para enviar respuestas de email
```
---
## Task Dependencies: Basic DAG
The `dependsOn` field turns `agent-tasks.json` into a minimal DAG (directed acyclic graph) without any external scheduler:
```json
[
{ "id": "task-run-tests", "title": "Run test suite", "status": "pending", "dependsOn": [] },
{ "id": "task-deploy-prod","title": "Deploy to prod", "status": "pending", "dependsOn": ["task-run-tests"] },
{ "id": "task-notify", "title": "Reply with result", "status": "pending", "dependsOn": ["task-deploy-prod"] }
]
```
The Executor resolves this automatically:
```javascript
function getNextTask(agentId) {
const queue = read(agentId);
const { tasks } = queue;
let dirty = false;
const pending = tasks
.filter(t => t.status === 'pending')
.sort((a, b) => (a.priority || 2) - (b.priority || 2));
let result = null;
for (const task of pending) {
if (!task.dependsOn?.length) { result = task; break; }
const deps = task.dependsOn.map(id => tasks.find(t => t.id === id));
const failedDep = deps.find(d => d?.status === 'failed');
if (failedDep) {
// Auto-skip tasks whose dependency failed β no manual intervention needed
task.status = 'skipped';
task.result = `Skipped: dependency "${failedDep.title}" failed`;
task.log.push({ ts: new Date().toISOString(), msg: task.result });
dirty = true;
continue;
}
if (deps.every(d => d?.status === 'done')) { result = task; break; }
}
if (dirty) write(agentId, queue);
return result;
}
```
If `task-run-tests` fails, `task-deploy-prod` and `task-notify` are automatically skipped. No cascade failure, no stuck queue.
---
## Idempotency: Handling Re-delivered Emails
Email systems re-deliver. Networks retry. The Planner must be idempotent:
```javascript
function addTask(agentId, newTask) {
const queue = read(agentId);
const existing = queue.tasks.find(t => t.threadId === newTask.threadId);
if (existing) {
if (existing.status === 'failed') {
// Explicit retry: reset to pending, increment retries counter
existing.status = 'pending';
existing.retries = (existing.retries || 0) + 1;
existing.log.push({ ts: new Date().toISOString(), msg: `Retry #${existing.retries}` });
write(agentId, queue);
}
// For any other status: silent no-op (dedup)
return queue;
}
queue.tasks.push({ retries: 0, log: [], dependsOn: [], ...newTask });
write(agentId, queue);
return queue;
}
```
The same `threadId` arriving twice results in one task. A `failed` task arriving again resets to `pending` with `retries++`. The `retries` field lets the Executor apply exponential backoff or hard-stop after N attempts.
---
## Per-Task Progress Logging
The `log[]` array makes task execution fully observable without any external logging infrastructure:
```json
{
"id": "task-deploy-prod",
"status": "done",
"log": [
{ "ts": "2026-05-31T10:00:01Z", "msg": "Iniciando: Deploy to prod" },
{ "ts": "2026-05-31T10:00:08Z", "msg": "SSH connected to 145.239.65.26" },
{ "ts": "2026-05-31T10:00:31Z", "msg": "docker pull done (847MB)" },
{ "ts": "2026-05-31T10:00:38Z", "msg": "Containers restarted, health check passed" }
],
"result": "Deployed v2.1.4 successfully"
}
```
Live monitoring: `watch -n1 'cat agent-tasks.json | jq ".tasks[0].log[-3:]"'`
No Grafana. No Datadog. The file is the dashboard.
---
## Crash Recovery
When the server restarts, tasks that were `in-progress` (mid-execution) need to be reset. A one-time recovery pass runs at startup:
```javascript
function recoverStaleTasks(agentId) {
const queue = read(agentId);
let recovered = 0;
for (const task of queue.tasks) {
if (task.status !== 'in-progress') continue;
task.status = 'pending';
task.retries = (task.retries || 0) + 1;
task.log.push({ ts: new Date().toISOString(), msg: 'Recovered from stale in-progress (server restart)' });
recovered++;
}
if (recovered > 0) write(agentId, queue);
return recovered;
}
// In startPoller():
function startPoller() {
const agents = listAgents().filter(a => a.enabled && a.ready);
for (const a of agents) taskQueue.recoverStaleTasks(a.id);
// ... start cron jobs
}
```
A crash mid-deploy becomes a retried deploy, not a lost task.
---
## Atomic Writes
The JSON file is the single source of truth. Corruption on a mid-write crash would break everything. The fix: write to a `.tmp` file, then `rename`:
```javascript
function write(agentId, data) {
const f = filePath(agentId);
const tmp = f + '.tmp';
fs.mkdirSync(path.dirname(f), { recursive: true });
fs.writeFileSync(tmp, JSON.stringify({ ...data, updatedAt: new Date().toISOString() }, null, 2));
fs.renameSync(tmp, f); // atomic on POSIX
}
```
`fs.renameSync` is atomic on POSIX filesystems: readers either see the old file or the new one, never a partial write.
---
## Timing and Concurrency
```
T+0s Email arrives
T+60s Poller tick detects new email β launches Scheduler
T+75s Scheduler reads email, writes tasks β exits
T+120s Executor tick fires β sees pending tasks β launches Executor
T+180s Executor completes tasks, replies to email β exits
T+240s Executor tick fires β no pending tasks β skips (noop)
```
Concurrency is controlled by in-memory flags that survive the process lifetime:
```javascript
// Only one Scheduler active per agent at a time
if (agentSessions.isActive(skillId, 'scheduler')) return null;
// Only one Executor active per agent at a time
if (agentSessions.isActive(skillId, 'executor')) return null;
```
Crucially: **the Scheduler and Executor can run simultaneously**. While the Executor is working on task #1, the Scheduler can plan tasks #2 and #3.
---
## Session Resumption
Both agents support `--resume`, which continues the same Claude conversation across invocations:
```javascript
const resumeClaudeId = sess.schedulerClaudeId || null;
const args = [
'--output-format', 'stream-json',
'--dangerously-skip-permissions',
'--model', resolvedModel,
];
if (resumeClaudeId) args.push('--resume', resumeClaudeId);
```
The Executor doesn't start cold. It already knows the project structure, conventions, and recent decisions. The more emails processed, the more efficient it becomes β without a database.
---
## Why Two Agents Instead of One?
| Concern | Single Agent | Planner-Executor |
|---------|-------------|-----------------|
| New email while executing | Missed until session ends | Scheduler handles it immediately |
| Task priority | FIFO only | Explicit priority 1-2-3 |
| Task dependencies | None | `dependsOn[]` with auto-skip on failure |
| Partial failure | Whole session fails | Task marked `failed`, next task continues |
| Crash mid-task | Task lost silently | Reset to `pending` on restart, `retries++` |
| Long-running tasks | Blocks everything | Executor runs async, Scheduler stays free |
| Debugging | One long session, hard to inspect | `agent-tasks.json` + `log[]` per task |
| Cost control | Hard to limit mid-session | Monthly spend limit checked before each launch |
| Re-delivered emails | May process twice | Dedup by `threadId`, idempotent |
---
## Spending Limits
Before launching either agent, the system checks monthly spend:
```javascript
if (agent.spendingLimitMonthly != null) {
const spent = await getMonthlySpend(skillId); // async β must be awaited
if (spent >= agent.spendingLimitMonthly) {
console.warn(`monthly limit $${agent.spendingLimitMonthly} reached β launch blocked`);
return null;
}
}
```
The Planner is cheap (reads email, writes JSON β a few cents). The Executor is where real work β and cost β happens. Checking before each launch gives per-session cost granularity.
---
## Complete Flow Diagram
```
Inbox (IMAP/Gmail)
β
β (polled every 60s)
βΌ
gmail-poller.js
β
ββ Is it a !command? β handle locally, reply, mark read
β
ββ Is sender authorized? β yes
β
βΌ
launchScheduler(agent, { from, subject, threadId })
β
β (Claude CLI, ROL: PLANIFICADOR)
β reads email thread via mail_client.py
β reads agent-tasks.json
β deduplicates by threadId
β appends tasks with priority + dependsOn
β writes agent-tasks.json (atomic)
ββ exits
(2 min later...)
tickExecutor()
β
ββ hasPendingReady(agentId)? β no β skip
β
ββ yes
β
βΌ
launchExecutor(agent)
β
β (Claude CLI, ROL: EJECUTOR)
β reads agent-tasks.json
β picks highest-priority pending task with deps satisfied
β marks in-progress, appends to log[]
β executes (code, email reply, API call, deploy...)
β marks done/failed with result + final log entry
β loops until queue empty or no more executable tasks
ββ exits
```
---
## Implementation in ClonAgent
The full implementation is open source:
| File | Role |
|------|------|
| [`server/lib/gmail-poller.js`](https://github.com/KikoCisBot/clonagent/blob/main/server/lib/gmail-poller.js) | Orchestrator: poller ticks, triggers Scheduler and Executor, crash recovery on startup |
| [`server/lib/relay-client.js`](https://github.com/KikoCisBot/clonagent/blob/main/server/lib/relay-client.js) | Launches Scheduler and Executor via Claude CLI with role-specific prompts |
| [`server/lib/task-queue.js`](https://github.com/KikoCisBot/clonagent/blob/main/server/lib/task-queue.js) | Atomic reads/writes of `agent-tasks.json`, idempotent `addTask`, dependency-aware `getNextTask`, crash recovery |
| [`server/lib/agent-sessions.js`](https://github.com/KikoCisBot/clonagent/blob/main/server/lib/agent-sessions.js) | Tracks active Scheduler/Executor sessions, prevents overlaps |
---
## Conclusion
The Planner-Executor pattern is a practical approach to building reliable AI agents that operate on asynchronous, real-world inputs like email.
The key insight is that **planning and execution are different cognitive tasks** that benefit from separation β not just conceptually, but as separate model invocations with different prompts, different time horizons, and different failure modes.
Production use has taught us additional lessons:
- **Atomic writes** (temp + rename) prevent file corruption that would break the whole agent
- **Idempotency by `threadId`** is mandatory β email systems re-deliver, networks retry
- **`log[]` per task** makes debugging possible without any external infrastructure
- **`dependsOn[]`** enables multi-step workflows without Airflow or Temporal
- **Crash recovery** at startup means a server restart doesn't lose work in progress
- **`await` the spend check** β async bugs here mean spending limits silently don't apply
A shared JSON file replaces complex message queue infrastructure. Claude CLI's `--resume` flag provides session continuity without a database. Explicit priority and dependency fields give the agent the ability to triage and sequence, just like a human would.
The result is an agent that feels less like a script and more like a colleague: it reads your email, decides what matters, plans the work, and executes β without blocking, without crashing, and without forgetting what it learned yesterday.
---
*Implementation: [ClonAgent](https://clonagent.utopiaia.com) β open source at [github.com/KikoCisBot/clonagent](https://github.com/KikoCisBot/clonagent)*
|