## Weekly signal Coverage note: this briefing is current through May 11, 2026; May 12 had not completed at publication. The week’s useful agent-collaboration signal was not more chatbot UX. It was infrastructure: multi-agent memory, delegation, cross-platform action, identity, and gateways.

The pattern is clear: agent collaboration is becoming a production architecture problem. Builders are being pushed to answer five questions: Which agent owns the task? How does it delegate? What shared state is allowed? Which actions can it take in enterprise systems? How do you audit every handoff?

## What changed 1. Anthropic moved Claude Managed Agents toward team-based agent work. On May 6, Anthropic launched dreaming in research preview and made outcomes, multiagent orchestration, and webhooks available to developers. The practical change is that a lead agent can split work across specialist agents with different prompts, models, and tools, while shared files and persistent events let the lead agent synthesize the result and trace what happened.

2. ServiceNow positioned itself as a governed action layer for third-party agents. Its May 5 Action Fabric announcement opened ServiceNow actions to agents built on Claude, Copilot, or custom stacks through a generally available MCP Server. The company says the server spans IT, HR, customer service, security, risk, compliance, and app development, which makes it relevant for coordinated multi-agent workflows across departments.

3. Agent governance became a collaboration prerequisite. ServiceNow also expanded AI Control Tower and deepened Microsoft Agent 365 integration, framing the problem as agent sprawl across platforms, tools, prompts, identities, and permissions. IBM’s Think 2026 announcements similarly emphasized agent orchestration, coordinated execution, and an operating model for managing agents at scale.

4. Security guidance caught up with agent swarms. CoSAI released agentic identity and security research after RSAC 2026, focused on machine-readable agent identity, safe delegation, intent-based authorization, and Agent Detection and Response. This matters because multi-agent systems multiply trust boundaries, not just API calls.

5. MCP plus A2A showed up in customer-service and gateway products. eGain launched Agentic Studio with multi-agent orchestration through MCP and A2A for end-to-end customer request resolution. The Linux Foundation agentgateway project also shipped v1.2.0-alpha.2 during the week; its README describes A2A support for capability discovery, modality negotiation, and task collaboration.

## What to do with it If you are building agent collaboration, do not start with a free-form group chat between agents. Start with explicit roles, narrow permissions, shared artifacts, and traceable handoffs.

For enterprise pilots, test one cross-system workflow where a lead agent delegates to two or three specialists, then measure latency, cost, error propagation, and review load. Add identity, policy, and observability before adding more agents.