Divide AI work into three steps
- GitHub, OpenAI, Anthropic, and Google are all now documenting the same AI workflow: give models the right context, let them use tools, then check the result instead of trusting one pass. - GitHub calls it a “three-part framework” for reliable workflows; OpenAI says GPT‑5.4 can “plan, execute, and verify” across long tasks; Google now lets Gemini pass context between tool calls. - The shift marks a move from prompt tricks to engineered systems with memory, tool access, and review loops. (github.blog)
Teams building AI systems are converging on a simple pattern: gather the right context, let the model use tools, then verify the output before anyone acts on it. (github.blog) (docs.langchain.com) The first step is context. GitHub says reliable workflows start with “context engineering,” and LangChain says missing context is more often the reason agents fail than weak model capability. (github.blog) (docs.langchain.com) In plain terms, context is the working file the model sees: instructions, message history, permissions, tool definitions, uploaded files, and prior tool results. LangChain breaks that into model context, tool context, and life-cycle context between steps. (docs.langchain.com) The second step is action. Anthropic’s Claude docs describe an “agentic loop” where the model requests a tool, code or infrastructure runs it, and the result comes back for the next step. (platform.claude.com) Google is pushing the same direction in Gemini. On March 17, 2026, Google said developers can combine Google Search, Google Maps, and custom functions in one request and keep the output of one tool available to the next. (blog.google) OpenAI’s latest model language now bakes the third step into the pitch. In its March 5, 2026 GPT‑5.4 launch, OpenAI said the model can “plan, execute, and verify tasks across long horizons” and search large tool ecosystems more efficiently. (openai.com) That verification step is the part that separates a chatbot answer from a production workflow. It can mean running tests, checking a source, matching a schema, comparing a number against a database, or sending a final pass to a human reviewer. (developers.openai.com) (platform.claude.com) Anthropic has been arguing for simpler building blocks rather than elaborate agent stacks. In its December 19, 2024 guide, the company said the strongest implementations it saw used “simple, composable patterns” instead of complex frameworks. (anthropic.com) That helps explain why the three-step pattern is spreading. A model does not need to know everything up front if the system can fetch the right facts, call the right service, and inspect the result before moving on. (anthropic.com) (developers.openai.com) The practical targets are the jobs that already look like checklists: repository automation, coding tasks with tests, internal search, report drafting, and service workflows that touch databases or application programming interfaces. GitHub says its agentic workflows run inside GitHub Actions for permissions, logging, auditing, and sandboxed execution. (github.blog) The new consensus is less “ask one giant prompt” and more “build a loop.” The model is becoming one part of a system that collects evidence, takes bounded actions, and proves the work before it ships. (github.blog) (openai.com)
