Weekly signal

This week tightened the infrastructure and market around multi-agent collaboration: open standards moved into neutral stewardship, major vendors shipped or previewed agent teams and proactive agents, and enterprise security pushed agent-aware detection into GA/early access. Key developments compress how agents will discover, share skills, coordinate work, and be monitored in production.

What changed

1. Anthropic donated its Agent Skills standard (the SKILL.md style packaging for reusable agent tasks) to the Agentic AI Foundation (AAIF) under the Linux Foundation — shifting control of a widely used interoperability primitive out of a single vendor and into an open-governance home.

2. The AAIF / Linux Foundation announced a set of new projects (Agent Skills among them) aimed at neutralizing agent interoperability pieces (Open Responses, OpenEnv, OpenSandbox), formalizing reuse and safety building blocks across vendors. That creates a clearer upgrade path for multi-agent workflows that must run across providers.

3. Anthropic opened a research preview of Claude Duet, a proactive agent app that runs background workflows, monitors connected apps, and checks in before taking actions — an example of agents acting autonomously and coordinating with humans and other services. The preview emphasizes background work, approvals, and shared plan usage.

4. Zoom announced a Unified Context Layer and ZoomMate Agent Teams that let a lead agent orchestrate specialist agents in parallel and consolidate outputs — a vendor-built pattern for agent collaboration inside enterprise workflows.

5. Security vendors and enterprise platforms responded: SentinelOne launched Singularity Agentic Detection & Response for discovering, monitoring and enforcing policy on agents at runtime; Oracle announced Claw-powered Fusion agentic HR apps and an agent marketplace that partners (e.g., Terillium) are populating — both signal that governance and runtime control are now product features.

What to do with it

• If you build agent workflows, adopt Agent Skills (or plan migration) now: the standard is under neutral governance, so author skills as reusable, versioned SKILL.md folders to maximize future portability and cross-vendor collaboration.

• Design approval/consent boundaries and shared-plan accounting into proactive agents (background work): Claude Duet showcases shared usage and explicit check-ins — test UX for false positives/urgent alerts before rolling agents into production.

• Add runtime observability and enforcement: treat agent actions as first-class events (discover, authorize, watch, block); pilot a runtime detection product or integrate enforcement hooks into your orchestration layer.

• Prototype lead-agent orchestration patterns (manager + specialists): ZoomMate and Oracle Claw show the model: small coordinator agents delegating to specialized skills reduces latency and clarifies responsibility. Build orchestration tests, ownership semantics, and failure modes.

• Engage with AAIF and vendor marketplaces: register skills, contribute security guidelines, and test cross-runner behavior in a staging MCP/Agent Skills environment to avoid vendor lock-in later.