Agentwatch
Sign InConsole
AI reliability

Measure AI agent reliability.
Before it becomes a problem.

Get StartedSign In

Every agent, one health score.

See drift the moment it starts — before your users do.

How it works

Wrap your agent. We watch the rest.

No thresholds to tune and no dashboards to configure. Three steps from install to the alert that matters.

  1. 01

    Install the SDK

    One wrapper around the function your agent already runs. Same arguments, same return value — about two minutes of work.

    terminal
    $ npm i agentwatch

    import { watch } from 'agentwatch'

    export const agent = watch(myAgent, {

    apiKey: process.env.AGENTWATCH_KEY,

    agentId: process.env.AGENTWATCH_AGENT_ID,

    })

    behaviour unchanged · returns the same value
  2. 02

    Your agent runs

    Every call is traced automatically: input, output, latency, tokens, cost. Nothing else to instrument, no dashboards to wire up.

    live traces4 of 12,847
    sess_8f2a41842ms1,284t
    sess_8f2a42910ms1,361t
    sess_8f2a431740ms2,208t
    sess_8f2a442260ms2,914t
    latency climbing · baseline 870ms
  3. 03

    Drift is caught automatically

    CUSUM and EWMA hold the baseline for you. The moment behaviour shifts, the alert goes to you — days before it reaches your users.

    drift detected2m ago
    support-bot72.6−13.3%

    CUSUM changepoint at day 6 · EWMA below baseline for 9 consecutive sessions.

    alert sent · #eng-oncall
Early access

AgentWatch is in early access.

Join the waitlist to get free access when we launch.

No spam. One email, when it's ready.