All AI sessions
Agent Atlas
Session 4 recap · Agents & Agentic Effort · Sun 18 Oct
One agent first; more only when the parts are independent. The one-page recap: the agent loop, tools and autonomy, subagents, agentic effort, four questions for any AI post, one myth, and a prompt to keep.
A prompt to keep
A plan first, the reads in parallel, every sentence checked against its file, and one result at the end.
Prepare the release note and the FAQ update from these files: the spec, the support tickets, the FAQ.
First show me a short plan and wait for my OK.
Read the three files separately, in parallel if you can.
Check every sentence you write against the file it comes from, and list what doesn't match.
Finish with one result: the release note, the FAQ changes, what's still open.Sources
Every claim on the sheet is checked against these.
- Building effective agentsAnthropic (opens in a new tab)
- A practical guide to building agentsOpenAI (opens in a new tab)
- Claude Cowork and chat are one ClaudeAnthropic (opens in a new tab)
- Get started with ChatGPT WorkOpenAI (opens in a new tab)
- Introducing advanced tool use (tool search)Anthropic (opens in a new tab)
- Connect tools with MCPClaude Code (opens in a new tab)
- Tool searchOpenAI (opens in a new tab)
- Permission modesClaude Code (opens in a new tab)
- Auto mode is now the default in Claude CodeAnthropic (opens in a new tab)
- Codex permission modesOpenAI (opens in a new tab)
- Codex approvals and securityOpenAI (opens in a new tab)
- SubagentsClaude Code (opens in a new tab)
- Codex subagentsOpenAI (opens in a new tab)
- Effective context engineering for AI agentsAnthropic (opens in a new tab)
- How we built our multi-agent research systemAnthropic (opens in a new tab)
- Multi-agent systems: when and how to use themAnthropic (opens in a new tab)
- Towards a science of scaling agent systemsGoogle Research (opens in a new tab)
- Multi-agents: what's actually workingCognition (opens in a new tab)
- Dynamic workflowsClaude Code (opens in a new tab)
- Codex long-running workOpenAI (opens in a new tab)
