Claude Managed Agents adds memory-store webhooks in July update

Claude Managed Agents can now trigger webhooks when an agent writes to long-term memory, per Releasebot. That closes the biggest gap in trusting long-running agents: before, you'd only find bad memory writes by auditing them later.

The update also lets developers set model effort per managed agent and add optional version checks. Webhook coverage now extends to environment events too, alongside session seeding with initial events and event-delta streaming for threads, per Releasebot.

Session seeding fixes a smaller but real annoyance. Developers have long replayed a fake conversation turn just to prime an agent with context before the real interaction starts. Seeding a session with initial events does the same job without leaving fake history in the transcript.

What builders should watch

Releasebot's writeup doesn't say whether memory-store webhooks fire before or after the write commits. That matters if you want to block a bad write instead of just logging it. I'd want that answered before wiring this into anything with real consequences, like a support bot that remembers customer preferences.

Reactive observability is a floor, not a ceiling. Whether Anthropic gives developers a way to veto a memory write before it lands is the next thing I'd watch for.