Spinning Up Agents

The power of Co-Work is explicit parallelism. And I need you to sit with that word — explicit — because it changes everything about how you think about prompting.

In Chat, you might write a prompt that says: "Analyze this website for SEO issues, UX improvements, accessibility problems, and content ideas." Claude will try to do all four sequentially in one pass. By the time it gets to content ideas, it has been working through a single context window that is already loaded with SEO analysis, UX recommendations, and accessibility findings. The quality degrades. Not because Claude is getting tired — it does not get tired — but because context is finite and attention is distributed.

In Co-Work, you say: "Spin up four sub-agents. Agent 1: audit SEO — meta tags, page speed, crawlability, backlink profile. Agent 2: analyze UX — navigation flow, mobile responsiveness, call-to-action placement. Agent 3: review accessibility — color contrast ratios, ARIA labels, keyboard navigation, screen reader compatibility. Agent 4: brainstorm content ideas — gaps in current content, keyword opportunities, competitor content analysis." Each agent gets its own fresh 200,000-token context window. Each is focused exclusively on its single task. Each can do independent web research without polluting the other agents' context.

1. 1. Provide the Initial Context Give Claude the starting material — a URL to browse, a document to analyze, a topic to research. The lead agent ingests this context and uses it to brief the sub-agents. In our website audit example, you would paste the URL and tell Claude to browse it first so it understands what it is working with.
2. 2. Request Parallel Sub-Agents Tell Claude explicitly to spin up N sub-agents with specific, independent tasks. Be precise about what each agent should focus on. Vague mandates produce vague results. "Agent 1: research SEO" is less effective than "Agent 1: audit meta tags, analyze page speed with Lighthouse methodology, assess crawlability, evaluate backlink profile."
3. 3. Monitor Individual Progress The Co-Work UI shows progress for each active agent. Watch for agents that stall or take significantly longer than their peers — this usually means the task needs to be decomposed further or the prompt was ambiguous. You can intervene with a specific agent without disrupting the others.
4. 4. Review and Synthesize Once all agents complete, the lead agent consolidates findings into a unified summary. Review it, then ask for specific deliverables: a prioritized action list, a slide deck, a detailed report stored in the working folder. The synthesis step is where the parallel work becomes a coherent output.

Here is a feature that sounds minor but changes your workflow once you internalize it: prompt queuing. While agents are still running, you can type and submit your next prompt. Co-Work will hold it in the queue and execute it the moment the current task finishes.

Why does this matter? Because you often know the next step before the current step finishes. You are running a research sprint and you already know you want the results compiled into a PowerPoint. You are generating a competitive analysis and you already know you want a one-page executive summary. Queue that prompt now. Co-Work chains them automatically.

This means you can structure multi-step workflows — research, then synthesis, then deliverable creation, then QA — submit the entire chain, and walk away. Come back to completed work instead of babysitting each stage. That is not a small thing. That is the difference between using an AI assistant and orchestrating an AI team.