The future

The future

The four D’s

One framework to take home, from Anthropic’s AI Fluency course: Delegation, Description, Discernment, Diligence. You did all four this week, whether you noticed or not:

  • Delegation: deciding what the dispatch agent does, and what stays human.
  • Description: CLAUDE.md, standing orders, behavior scenarios.
  • Discernment: PR reviews, evals, reading what the loop actually did.
  • Diligence: typed tools, deterministic checks, tick and spend caps.

Where to go next

Keep building. For the theory behind what you built, Anthropic Academy’s free courses come with certificates: Claude Code in Action, Introduction to agent skills, and the MCP courses. Directions to explore, roughly in order of how close they sit to this course:

  • Build an MCP server for a tool you already use. The connective tissue every agent after this one will need.
  • Other harnesses: OpenCode, Codex, Cursor, Amp. Same loop, different trade-offs.
  • Alternative models inside the harness, and local models, for work that cannot leave the building.
  • Token use optimisation. You saw /context, /compact, and prompt caching; there is a whole discipline in there.
  • Agentic workflow automation and multi-agent systems, for when one loop stops being enough.

Teach someone else. Run an internal hackathon; the demo-and-issue-filing format works at any scale. AI Fluency is the course for the colleagues you will show this to.

Other things to explore

Claude Code features the course never needed, worth a try on your own time:

  • /voice dictates your prompts: talking through a task is often faster than typing it, and rambling out loud surfaces requirements you would never bother to write down. There is audio output too - try /radio for Claude FM. (Voice in Claude Code)
  • /advisor pairs your working model with a smarter one, consulted only at key moments - for a plan or a course correction. Judgement where it counts, cheap tokens everywhere else. (Claude Code Advisor Command Explained in 3 Minutes)
  • Claude Code with other models: GLM, DeepSeek, and Kimi serve Anthropic-compatible endpoints, so pointing Claude Code’s environment variables at their base URL and key swaps the brain inside the same loop and tools - the most direct demonstration of the harness/model split. (Z.ai GLM setup; DeepSeek’s equivalent)
  • Headless mode runs the whole agent as one command that prints a result and exits: claude -p and opencode run. This is the shape a scheduler invokes for true unattended operation. (Headless docs)

Claude Code on the web

Claude Code running in a managed VM at claude.ai/code: it clones your repo and keeps working after you disconnect, so a closed laptop lid no longer ends the session. (Docs)

Launch a background run

One long-running task, handed to the web while you keep working locally:

  • Connect your GitHub repo at claude.ai/code and start a session; the managed VM keeps working after you close the tab.
  • Hand it one bounded goal with /goal, so the checker keeps it honest between turns. A refactoring pass - tests, structure, dead code, docstrings - is a good first candidate: useful, and hard to break anything.
  • Keep it off your branch: it works on its own branch and comes home as a PR you review.
  • The goal rules apply even to a two-hour run: a target it cannot game, constraints with checking instruments, and caps. Here the caps are /goal’s checker plus your account’s usage limits - know both before you start.
  • Get it running, then forget about it. That is the point.

Further reading

Who to follow

You just spent two days building an agent that reads contacts and acts on them. These are feeds. Build the loop. Stay the engineer.