[rparundekar]

What this does

Adds a "Workflow or agent?" section to the root README, following the cookbook guide When to use a workflow, and when an agent. It names the three cases where an agent earns its place:

• someone at the screen asking and replying;
• steps nobody can write down, but a result code can check;
• a workflow handing off an input it wasn't built for.

Each agentic recipe gets a Shape column saying which case it shows and what checks it.

• AGENTS.md: describes the new layout. A workflow folder may have an agent.py that imports recipe.py and keeps its extra data in data/agent/. Its anthropic dependency goes in an agent dependency group, and the workflow's results must not change.
• CI: smoke-tests agent.py --help wherever one exists.
• Screening recipe: agentic/screen_applicants_through_mcp_server gets an honest note. By the test above the job is really a workflow, and it's kept as an agent to show masking enforced in an MCP server.

The four agents beside workflows land in their own PRs. Merge this one first.

🤖 Generated with Claude Code

[rparundekar]

Self-review

I reviewed this PR inline: 54 lines across four files, one of them a single CI step. A fan-out of review agents would have cost more than reading it. The security pass found nothing: the new CI step runs agent.py --help from the recipe's own folder, with no secrets in the environment.

Important (2)

• README.md:63: the link to the workflow-vs-agent post returned 404, because the post isn't published yet. Fixed: removed it. The section now lists the four workflows that carry an agent.py. The link comes back with the blog update.
• AGENTS.md:25: said "CI never runs agent.py", but this PR adds a CI step that runs agent.py --help. Fixed: it now says CI runs only --help, never a live agent run.

Fixed in 957a4ce.