adk-task-scheduler

adk-task-scheduler

Auto-scheduling for Google ADK agents.
Bake cron, interval, or condition-based triggers directly into your agent definitions so they self-wake — no separate orchestrator, no cloud scheduler, no code changes to ADK.

---

Why

ADK agents are purely invocation-driven. Out of the box an agent only runs when a user or external system sends a message. adk-task-scheduler adds a self-wake capability that coexists with the standard ad-hoc POST /run interface:

Without this library                  With this library
─────────────────────────────         ──────────────────────────────────────────
User/system → POST /run → agent       User/system → POST /run → agent (unchanged)
                                      APScheduler tick → agent (new)

---

How it works

get_fast_api_app (the ADK function that powers adk api_server) accepts a lifespan= parameter. This library builds an APScheduler AsyncIOScheduler lifespan and passes it in — no monkey-patching, no ADK fork required. ADK's own runner_dict handles all ad-hoc calls; the scheduler maintains a separate RunnerPool so the two paths never interfere.

┌──────────────────────────────────────────────────────────────┐
│  FastAPI app  (returned by build_scheduled_app)              │
│                                                              │
│  ┌─────────────────────────┐  ┌──────────────────────────┐  │
│  │  ADK runner_dict        │  │  Scheduler RunnerPool    │  │
│  │  POST /run              │  │  cron / interval /       │  │
│  │  POST /run_sse          │  │  condition triggers      │  │
│  │  WebSocket /run_live    │  │                          │  │
│  │  GET/POST /a2a/...      │  │  Separate Runner per     │  │
│  └─────────────────────────┘  │  app_name — no overlap   │  │
│                                └──────────────────────────┘  │
│        ADK lifespan ── scheduler lifespan (composed)         │
└──────────────────────────────────────────────────────────────┘

---

Installation

pip install adk-task-scheduler

Requirements: Python ≥ 3.10, google-adk ≥ 1.7.0

---

Quick start

1. Attach a schedule to root_agent

Note: Python's @decorator syntax only applies to def/class statements, not variable assignments. Use with_schedule() (the recommended API) or the scheduled()(agent) call form.

# agents/ticker_agent/agent.py
from google.adk.agents import Agent
from adk_task_scheduler import with_schedule

root_agent = with_schedule(
    Agent(
        name="ticker_agent",
        model="gemini-2.0-flash",
        instruction="""
        You are a periodic monitoring agent.

        When the user message is exactly '__tick__', you are being invoked by
        the scheduler (not a real user). Run your monitoring routine and report
        status. Do NOT ask clarifying questions.

        For any other message, respond normally as a helpful assistant.
        """,
    ),
    interval_seconds=60,
    trigger_text="__tick__",
    on_response=lambda text: print(f"[monitor] {text}"),
)

2. Build the app

# main.py
from adk_task_scheduler import build_scheduled_app

app = build_scheduled_app(
    agents_dir="./agents",
    auto_discover=True,   # picks up with_schedule() agents automatically
    web=False,
    a2a=True,
)

uvicorn main:app --host 0.0.0.0 --port 8000

The agent now:

• Responds to ad-hoc POST /run calls (standard ADK — unchanged).
• Self-wakes every 60 seconds and logs its response.

---

Examples

Cron-triggered daily briefing agent

# agents/briefing_agent/agent.py
import logging
from google.adk.agents import Agent
from adk_task_scheduler import with_schedule

logger = logging.getLogger(__name__)

root_agent = with_schedule(
    Agent(
        name="briefing_agent",
        model="gemini-2.0-flash",
        instruction="""
        You are a morning briefing assistant.

        When the user message is '__morning__', produce a concise daily briefing:
        - Key tasks for the day (you may fabricate plausible examples)
        - Weather summary for London
        - One motivational note

        For other messages, respond normally.
        """,
    ),
    cron="0 8 * * 1-5",          # 08:00 every weekday
    trigger_text="__morning__",
    user_id="briefing-system",
    session_service_uri="sqlite:///./briefing.db",   # persist sessions
    on_response=lambda text: logger.info("Daily briefing:\n%s", text),
    on_error=lambda exc: logger.error("Briefing failed: %s", exc),
)

---

Condition-triggered market monitor

# agents/market_monitor/agent.py
import logging
from datetime import datetime, timezone
from google.adk.agents import Agent
from adk_task_scheduler import with_schedule

logger = logging.getLogger(__name__)


def is_market_open() -> bool:
    """True during NYSE trading hours Mon-Fri 14:30-21:00 UTC."""
    now = datetime.now(timezone.utc)
    if now.weekday() >= 5:          # weekend
        return False
    return 14 <= now.hour < 21


root_agent = with_schedule(
    Agent(
        name="market_monitor",
        model="gemini-2.0-flash",
        instruction="""
        You are a market monitoring agent.

        When the user message is '__market_check__', analyse current market
        conditions (simulate with plausible data) and flag any anomalies.
        Keep the response under 3 bullet points.

        For other messages respond normally.
        """,
    ),
    condition=is_market_open,          # evaluated every 5 minutes
    condition_poll_interval=300,
    trigger_text="__market_check__",
    max_concurrent_runs=1,             # never overlap
    on_response=lambda text: logger.info("[market] %s", text),
)

---

Multi-agent: mix scheduled and ad-hoc in one app

agents/
├── briefing_agent/
│   ├── __init__.py
│   └── agent.py        ← with_schedule(cron="0 8 * * 1-5")
├── market_monitor/
│   ├── __init__.py
│   └── agent.py        ← with_schedule(condition=is_market_open)
└── assistant/
    ├── __init__.py
    └── agent.py        ← plain Agent, no schedule (ad-hoc only)

# main.py
from adk_task_scheduler import build_scheduled_app

app = build_scheduled_app(
    agents_dir="./agents",
    auto_discover=True,          # discovers all three agents
    web=False,
    a2a=True,
    session_service_uri="sqlite:///./sessions.db",
)

auto_discover=True (default) scans agents_dir for any root_agent that carries a schedule config. Agents without a schedule are registered in ADK's router as usual and remain fully available for ad-hoc calls.

---

Explicit ScheduleConfig (wiring outside agent.py)

Useful when the schedule is environment-specific (e.g. different cron for staging vs. production) and you don't want it hard-coded in the agent file:

# main.py
import os
from adk_task_scheduler import ScheduleConfig, build_scheduled_app
from agents.reporter.agent import root_agent   # plain Agent, no schedule

cfg = ScheduleConfig(
    agent=root_agent,
    cron=os.environ["REPORT_CRON"],            # e.g. "0 6 * * *"
    trigger_text="__generate_report__",
    user_id="report-system",
    session_service_uri=os.environ["DB_URI"],
    on_response=lambda text: publish_to_slack(text),
    on_error=lambda exc: alert_oncall(exc),
    max_concurrent_runs=1,
    misfire_grace_time=120,
)

app = build_scheduled_app(
    agents_dir="./agents",
    schedules=[cfg],
    auto_discover=False,   # opt out of scanning when configs are explicit
    web=False,
    a2a=True,
)

---

Trigger types

---

API reference

with_schedule(agent, *, ...) ← recommended

from adk_task_scheduler import with_schedule

root_agent = with_schedule(
    Agent(name="my_agent", ...),
    cron="0 * * * *",
    trigger_text="__tick__",
    session_service_uri="sqlite:///./sessions.db",
    on_response=lambda text: print(text),
    on_error=lambda exc: print(f"Error: {exc}"),
)

scheduled(**kwargs) → Callable[[BaseAgent], BaseAgent]

Equivalent to with_schedule but curried:

from adk_task_scheduler import scheduled

root_agent = scheduled(interval_seconds=30)(Agent(...))

All parameters identical to with_schedule.

build_scheduled_app(*, agents_dir, ...) → FastAPI

Drop-in for get_fast_api_app. All get_fast_api_app keyword arguments (a2a, session_service_uri, allow_origins, trace_to_cloud, etc.) are forwarded verbatim.

ScheduleConfig

Dataclass holding all scheduling metadata. Construct directly when you need environment-specific configuration that shouldn't live in agent.py.

---

Session strategy

Each scheduled invocation creates a fresh session by default, so runs are fully isolated. To retain state across ticks (e.g. accumulate context over time), pass a session_service_uri pointing to a persistent store and manage session IDs yourself via extra_state or the on_response callback.

ScheduleConfig(
    agent=root_agent,
    interval_seconds=3600,
    session_service_uri="sqlite:///./scheduler.db",
    extra_state={"environment": "production", "tenant": "acme"},
)

---

Development

git clone https://github.com/STHITAPRAJNAS/adk-task-scheduler.git
cd adk-task-scheduler

python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"

# Run tests
pytest

# Lint
ruff check .

Running the example

cd example
pip install -e "..[dev]"   # install from parent
uvicorn main:app --reload --log-level info

The ticker_agent will fire every 30 seconds and log its response. You can also call it ad-hoc:

curl -s -X POST http://localhost:8000/run \
  -H "Content-Type: application/json" \
  -d '{
    "app_name": "ticker_agent",
    "user_id": "me",
    "session_id": "test-1",
    "new_message": {"role": "user", "parts": [{"text": "What time is it?"}]}
  }' | python -m json.tool

---

Releasing

This project uses PyPI Trusted Publishing (OIDC) — no API token needed.

Steps:

1. Bump the version in pyproject.toml and adk_task_scheduler/__init__.py.
2. Add a section to CHANGELOG.md.
3. Commit and push to main.
4. Tag the release:

   git tag v0.2.0
   git push origin v0.2.0
   

5. The publish workflow triggers automatically on v*.*.* tags, runs the full test suite, then uploads to PyPI via pypa/gh-action-pypi-publish.

One-time setup (first release only):
Add a Trusted Publisher entry on pypi.org/manage/account/publishing/ pointing to STHITAPRAJNAS/adk-task-scheduler, workflow publish.yml, and environment pypi.

---

License

Apache 2.0 — see LICENSE.