How to Orchestrate Multi-Agents in Claude?

11 Просмотры

Спасибо! Поделитесь с друзьями!

Вам не понравилось видео. Спасибо за то что поделились своим мнением!

Издатель Jun 1, 2026

Last Thursday I presented "How to Orchestrate Multi-Agents in Claude?" with a live demo on a real Salesforce delivery use case.

Three things I wanted the room to walk away with.
- Multi-agent changes the economics of Salesforce delivery. Instead of one generalist agent doing schema, automation, Apex, and documentation, I set up a coordinator that routes the work to four specialists: data-architect, flow-builder, apex-tester, solution-writer. One user story in. Object XML, Flow XML, Apex test class, and a client-facing implementation doc out.

- The quality lift is real. So is the cost. Anthropic's engineering team published a benchmark showing multi-agent systems beat single agents by 90.2%. At 15x the tokens. The question isn't "more agents = better." It's: where does the outcome justify the bill?

- Sub-agents and Agent Teams solve different problems.
Sub-agents run synchronously. One at a time. Parent evaluates each result before the next step. Best when work has dependencies; like a Salesforce build, where the flow needs the schema and the test needs the flow.

Agent Teams run in parallel through a shared task list. Best when streams are independent and outputs merge.

The demo ran end-to-end on platform.claude.com against a live Salesforce sandbox via the official Hosted MCP server. Coordinator decomposes the story, four session threads spawn, each worker stays in isolated context, coordinator synthesises.

Thank you to Dominik Fretz, Bojan Zivic, Claude Community Australia, Anthropic, Amazon Web Services (AWS) for organising and hosting!

#Salesforce #Agentforce #MCP