Commitment-as-a-Service (CaaS)

Others persist computation. CaaS persists responsibility.

Promises kept — even when the promiser is gone.

---

The Problem

User:  "Remind me to review the PR in 10 minutes."
Agent: "Got it! I'll remind you."

     ↓ 5 minutes later

╔═══════════════════════════════════╗
║  SESSION TIMEOUT - AGENT DIED     ║
╚═══════════════════════════════════╝

     ↓ 10 minutes later

     ...nothing happens.
     Promise broken.

AI agents are stateless, ephemeral, and can't schedule. You're binding a future obligation to something that cannot exist in the future.

---

The Solution

Agent authors  ───▶  CaaS owns  ───▶  Webhook fires
    intent           obligation        (guaranteed)

  [ephemeral]        [durable]        [delivered]
  [can die]          [survives]       [retried]

curl -X POST /v1/commitments -d '{
  "trigger_type": "time_after",
  "trigger_config": {"timestamp": "2024-01-15T10:30:00Z"},
  "obligation_config": {"target": "https://app.com/webhook", "payload": {...}}
}'

The agent dies. The promise lives. The webhook fires.

---

CaaS vs Temporal

Yes, Temporal can technically do this. The difference is the abstraction.

	Temporal	CaaS
Unit of work	Workflow execution	Commitment
To remind in 10 min	Define workflow → Deploy workers → Start execution	POST JSON
What persists	Execution history	Obligation record
Mental model	"Run this code durably"	"Keep this promise"

# Temporal: define workflow, deploy workers, start execution
@workflow.defn
class ReminderWorkflow:
    @workflow.run
    async def run(self, target, message, delay):
        await asyncio.sleep(delay)
        await workflow.execute_activity(send_webhook, target, message)

client = await Client.connect("temporal:7233")
await client.start_workflow(ReminderWorkflow.run, ...)

# CaaS: POST and leave
curl -X POST /v1/commitments -d '{"trigger_type": "time_after", ...}'

When Temporal wins: Complex control flow, human-in-the-loop, queryable state, existing infrastructure.

When CaaS wins: Simple obligations, no workflow definitions, agents that just want to POST and leave.

---

What Only CaaS Can Do

Novel obligations without pre-deployed code:

# Agent composes this on the fly — no one pre-wrote a workflow for it
curl -X POST /v1/commitments -d '{
  "trigger_type": "compound",
  "trigger_config": {
    "all": [
      {"event_type": "price.update", "filters": {"symbol": "ETH", "price_gte": 5000}},
      {"event_type": "time.weekday"},
      {"event_type": "user.presence", "filters": {"status": "online"}}
    ]
  },
  "obligation_config": {"target": "https://app.com/notify", "payload": {"alert": "buy"}}
}'

	Temporal	CaaS
New trigger combination	Write code, deploy workers	POST JSON
Agent capability	Invoke pre-defined workflows	Author novel obligations

Temporal requires someone to have written the workflow. CaaS lets the agent be the workflow author.

---

Quick Start

git clone https://github.com/lulzx/caas.git && cd caas
docker-compose up -d
docker-compose exec app mix ecto.migrate
curl http://localhost:4000/health

---

API

Create Commitment

# Time-based
curl -X POST /v1/commitments -H "Authorization: Bearer $KEY" -d '{
  "trigger_type": "time_after",
  "trigger_config": {"timestamp": "2024-01-15T10:30:00Z"},
  "obligation_config": {"target": "https://app.com/hook", "payload": {"msg": "hi"}}
}'

# Event-based
curl -X POST /v1/commitments -d '{
  "trigger_type": "webhook",
  "trigger_config": {"event_type": "payment.completed", "filters": {"amount_gte": 100}},
  "obligation_config": {"target": "https://app.com/hook", "payload": {"action": "receipt"}}
}'

# Tool completion
curl -X POST /v1/commitments -d '{
  "trigger_type": "tool_complete",
  "trigger_config": {"tool_id": "code_reviewer", "outcome": "success"},
  "obligation_config": {"target": "https://app.com/hook", "payload": {"next": "merge"}}
}'

Send Events

curl -X POST /v1/events -d '{
  "event_type": "tool_complete",
  "event_key": "code_reviewer",
  "payload": {"tool_id": "code_reviewer", "outcome": "success"}
}'

Endpoints

Method	Path	Description
POST	/v1/commitments	Create commitment
GET	/v1/commitments	List commitments
GET	/v1/commitments/:id	Get commitment
DELETE	/v1/commitments/:id	Cancel commitment
POST	/v1/commitments/batch	Batch create
POST	/v1/commitments/batch/cancel	Batch cancel
POST	/v1/events	Send event
GET	/health	Health check
GET	/metrics	Prometheus metrics

---

Patterns

Saga — Chain commitments for multi-step transactions:

[Payment] ──▶ [Inventory] ──▶ [Shipping] ──▶ [Confirm]

Escalation — Time-based SLA monitoring:

0min ──▶ 5min ──▶ 15min ──▶ 30min
         Assign   Escalate   PAGE

Dead Man's Switch — Alert if no heartbeat:

[Heartbeat] resets timer. No heartbeat? ──▶ ALERT

Fan-out/Fan-in — Coordinate parallel agents:

Orchestrator spawns [A1] [A2] [A3] ──▶ All complete ──▶ Next step

---

Architecture

Each commitment is an isolated Erlang process (OTP GenServer):

┌─────────────────────────────────────────────────────┐
│              CommitmentSupervisor                   │
├─────────────────────────────────────────────────────┤
│  [Server]  [Server]  [Server]  [Server]  ...        │
│   C-001     C-002     C-003     C-004               │
└─────────────────────────────────────────────────────┘

• Isolation: One failure doesn't affect others
• Fault tolerance: Crashed processes restart automatically
• Concurrency: Thousands of commitments in parallel

Commitment Lifecycle

        ┌──────────┐
        │ pending  │
        └────┬─────┘
   ┌─────────┼─────────┐
   ▼         ▼         ▼
cancelled triggered  expired
             │
       ┌─────┴─────┐
       ▼           ▼
   fulfilled    failed

Retry

%{"max_attempts" => 5, "backoff" => "exponential", "initial_delay_seconds" => 5}
# Attempts: immediate → 5s → 25s → 125s → 625s

---

Configuration

Variable	Default
DATABASE_URL	-
SECRET_KEY_BASE	-
PORT	4000
POOL_SIZE	10

---

Deployment

# Fly.io
fly launch && fly secrets set SECRET_KEY_BASE=$(mix phx.gen.secret) && fly deploy

# Docker
docker build -t caas . && docker run -p 4000:4000 -e DATABASE_URL=... -e SECRET_KEY_BASE=... caas

---

Demos

./demo_concept.sh   # The "aha moment"
./demo_pipeline.sh  # AI code review pipeline
./demo_showcase.sh  # 5 patterns: saga, escalation, heartbeat, fan-out, batch

---

The Bottom Line

	Guarantee
Temporal	"If this workflow is running, it will finish."
CaaS	"Someone promised this — therefore it will happen."

Others persist computation. CaaS persists responsibility.

Agents come and go. Commitments persist.

---

MIT License