Watcher

A closer look at Watcher

Watcher by Apollo Research is a runtime monitoring and control tool for AI agents, built for production engineering teams at agent-building companies, quant trading firms, and multinationals. It deploys one set of policies across an organization, checks agent actions before they run, and blocks those crossing a risk threshold, with preconfigured policies for acting beyond user intent, prompt injection, privilege escalation, and 47 more risks. Sessions are reviewed for incidents, findings go to tools such as Splunk, Datadog, Grafana, Elastic, and Slack, and Watcher can be installed self-serve or self-hosted.

Editorial guidance based on published documentation. The workflow below is a suggested starting point, not a measured product result.

When to choose Watcher

• Engineering and security teams running fleets of coding agents, since Watcher monitors agent tokens at scale and gives one set of policies with visibility across the team's agents.
• Teams that need risky actions stopped before they run, since Watcher checks agent actions before they run and blocks those crossing a risk threshold, including prompt injection and privilege escalation.

When to choose something else

• Self-serve users who need a harness other than Claude Code, Codex, or Copilot CLI, since the self-serve version integrates directly with those three, with 50+ harnesses on enterprise.
• Organizations that need self-hosting without a sales conversation, since self-hosting is discussed over a booked call on the page read.

A practical starting workflow

1. Install Watcher self-serve with the documented install command, or discuss self-hosting over a booked call.
2. Deploy Apollo's preconfigured policies in one click, or add custom rules for the environment.
3. Let Watcher check agent actions before they run and block those crossing the risk threshold in real time.
4. Review sessions, investigate findings, and route them to Splunk, Datadog, Grafana, Elastic, Slack, or other team tools.

Check before committing

• Recall and false-positive figures are measured on a synthetic attack dataset, while cost and latency are estimated from real agent traffic, so they are not guarantees for a new deployment.
• The install command on the page connects to Apollo's managed monitoring platform in Western Europe, so data-residency needs should be checked before installing.
• Self-hosting is available, but the page says to book a call to discuss deployment, so it is not a self-serve path on the page read.

Common questions

Compare approaches for this task

• Waxell — Engineering and security teams running many AI agents, since Waxell gives one governance plane with visibility into every agent, model call, and agentic workflow.; Teams that need governed agent deployment, since Waxell offers governance policy categories, MCP gateway governance, and managed runtime on higher tiers.
• Claude — Working through source material, developing structured drafts, and iterating on code or documents.