Ag2 Observers And Alerts

Observers, watches, and alerts

When to use

• Observability — log model responses, tool calls, token usage.
• Runtime safety — block dangerous tool arguments, halt the agent.
• Reactive metrics — fire on every Nth response, or every M seconds.
• Loop / repetition detection — catch infinite tool-call loops.
• Stateful monitoring — anything that needs to remember prior events to decide what to do next.

Two observer shapes

Both are stream subscribers under the hood — registered on the agent rather than directly on the stream.

60-second recipe — @observer

from ag2 import Agent, observer
from ag2.config import OpenAIConfig
from ag2.events import ModelResponse

@observer(ModelResponse)
async def log_response(event: ModelResponse) -> None:
    print(f"Model said: {event.content}")

agent = Agent(
    "assistant",
    config=OpenAIConfig(model="gpt-4o-mini"),
    observers=[log_response],
)

Or attach after construction with @agent.observer(...). Per-call observers also supported (agent.ask("...", observers=[...])).

Observer callbacks support full dependency injection (Context, Inject, Variable, Depends). Filter by event type, multiple types (ModelRequest | ModelResponse), or field value (ToolCallEvent.name == "search"). Use interrupt=True to modify or suppress events before regular subscribers see them.

Built-in stateful observers

from ag2 import Agent
from ag2.observers import LoopDetector, TokenMonitor

agent = Agent(
    "assistant",
    config=config,
    observers=[
        TokenMonitor(warn_threshold=50_000, alert_threshold=100_000),
        LoopDetector(window_size=10, repeat_threshold=3),
    ],
)

• TokenMonitor — tracks cumulative tokens across ModelResponse and TaskCompleted. Emits WARNING / CRITICAL ObserverAlerts as thresholds are crossed. Read state via monitor.total_tokens.
• LoopDetector — sliding window of recent tool calls. Emits a WARNING alert when repeat_threshold consecutive identical calls are seen.

Custom BaseObserver

A BaseObserver pairs a Watch (when to fire) with a process() method (what to do):

from ag2 import Context
from ag2.observers import BaseObserver
from ag2.watch import CadenceWatch
from ag2.events import BaseEvent, ModelResponse
from ag2.events.alert import ObserverAlert, Severity

class AvgCompletionObserver(BaseObserver):
    """Every N responses, emit an INFO alert with avg completion-token count."""

    def __init__(self, window: int = 5) -> None:
        super().__init__("avg-completion", watch=CadenceWatch(n=window, condition=ModelResponse))
        self._window = window

    async def process(self, events: list[BaseEvent], ctx: Context) -> ObserverAlert | None:
        tokens = [e.usage.completion_tokens for e in events if isinstance(e, ModelResponse) and e.usage]
        if not tokens:
            return None
        return ObserverAlert(
            source=self.name,
            severity=Severity.INFO,
            message=f"Avg completion tokens over last {self._window}: {sum(tokens) / len(tokens):.0f}",
        )

If process() returns an ObserverAlert, the base class emits it onto the stream. You can also send events manually via await ctx.send(...).

Watch primitives — picking when to fire

All importable from ag2.watch. Callback signature is uniform: async def cb(events: list[BaseEvent], ctx: Context) -> None. Time-driven watches pass events=[].

ObserverAlert — the alert type

from ag2.events.alert import ObserverAlert, Severity

ObserverAlert(
    source="my-observer",
    severity=Severity.WARNING,    # INFO, WARNING, CRITICAL, FATAL
    message="What happened",
)

Important: ObserverAlert is on the stream and persisted in history, but the default provider mappers do not render it back to the LLM. To make the agent see alerts, add AlertPolicy() to assembly=[...]:

from ag2.policies import AlertPolicy
agent = Agent("assistant", config=config, assembly=[AlertPolicy()])

FATAL alerts → HaltEvent → short-circuit

AlertPolicy does two things on Severity.FATAL:

1. Emits a HaltEvent on the stream.
2. Appends a halt notice to the system prompt.

When assembly=[...] is non-empty, the harness automatically wires _HaltCheckMiddleware which sees the HaltEvent and short-circuits the next LLM call with a synthetic HALTED: ... response.

from ag2 import Context
from ag2.observers import BaseObserver
from ag2.events import BaseEvent, ToolCallEvent
from ag2.events.alert import HaltEvent, ObserverAlert, Severity
from ag2.policies import AlertPolicy
from ag2.watch import EventWatch

class PathGuardian(BaseObserver):
    def __init__(self) -> None:
        super().__init__("path-guardian", watch=EventWatch(ToolCallEvent))

    async def process(self, events: list[BaseEvent], ctx: Context) -> ObserverAlert | None:
        for event in events:
            if not isinstance(event, ToolCallEvent) or event.name != "write_file":
                continue
            if "/etc/" in event.arguments or "/usr/" in event.arguments:
                return ObserverAlert(
                    source=self.name,
                    severity=Severity.FATAL,
                    message=f"blocked dangerous write: {event.arguments}",
                )
        return None

agent = Agent(
    "safe-shell",
    prompt="...",
    config=config,
    tools=[write_file],
    observers=[PathGuardian()],
    assembly=[AlertPolicy()],   # routes FATAL → HaltEvent
)

The first dangerous tool call triggers FATAL → halt; the agent's next ask is short-circuited. Full runnable demo: assets/safety_guard.py.

Subscribing to alerts and halts from outside

from ag2 import MemoryStream
from ag2.events.alert import HaltEvent, ObserverAlert

stream = MemoryStream()
stream.where(ObserverAlert).subscribe(lambda e: print(f"[{e.severity}] {e.source}: {e.message}"))
stream.where(HaltEvent).subscribe(lambda e: print(f"HALT: {e.reason}"))
await agent.ask("...", stream=stream)

Observers vs Middleware vs Stream subscribers

Going deeper

• assets/token_watchdog.py — three observers (TokenMonitor, LoopDetector, custom AlertConsole) on one agent. Mirrors code_examples/04.
• assets/safety_guard.py — PathGuardian → FATAL → AlertPolicy → HaltEvent → short-circuit. Mirrors code_examples/08.
• Source docs:
  • website/docs/user-guide/advanced/observers.mdx — @observer, BaseObserver, registration, built-ins, ObserverAlert.
  • website/docs/user-guide/advanced/watches.mdx — every Watch primitive, composition rules.
  • website/docs/user-guide/advanced/stream.mdx — Stream API, where, subscribe, interrupters, RedisStream.
  • website/docs/user-guide/advanced/assembly.mdx — AlertPolicy ordering and dedup.

Common pitfalls

• Alerts not reaching the model — ObserverAlert events are on the stream but invisible to the LLM by default. Add AlertPolicy() to assembly=[...].
• FATAL not halting — AlertPolicy is what creates HaltEvent. Without assembly=[..., AlertPolicy(), ...] (or any non-empty assembly chain enabling _HaltCheckMiddleware), nothing halts.
• Sharing one AlertPolicy() across agents — dedup state lives on the instance. Give each agent its own.
• Watch callback assumes events is non-empty — for time-driven watches (DelayWatch, IntervalWatch, CronWatch), events is always [].
• Forgetting process() is async — BaseObserver.process must be async def.
• Subscribing with subscribe(fn) when you wanted subscribe() decorator — both work; the bare-call form is stream.subscribe(fn), the decorator form is @stream.subscribe() (with parens).
• CadenceWatch with no n and no max_wait — raises ValueError; at least one is required.