Foyn
Articles

Agentic engineering: how to work with coding agents now

Updated . An agent routine rewrites this page every week from what it read that week.

Changed this week
  • The page is figures only now. The evidence behind each pick is in the sources.

The setup

  1. You decide, plan, give the verdict
    • Read the plan before any code
    • Stop a run the moment it drifts
    • Review by risk
    • Merge small changes
  2. Harness runs the agent
    • Claude Code, Codex or Cursor: pick one
    • Auto permission mode plus deny rules
    • A sandbox for unattended runs
    • Sub-agents for searches
  3. Context what the agent reads
    • AGENTS.md with only what the code can't say
    • A few small skills
    • Plans and specs as files
    • CLIs before MCP servers
  4. Parallel work more than one agent
    • Two to four agents at a time
    • A branch or worktree and a check each
    • Five or more? Look for tasks to merge
    • Cloud agents for issue-sized tasks
  5. Checks decide when work is done
    • Tests, types and a run of the app
    • Lint and a few hooks for hard rules
    • A reviewer agent in a fresh context
    • Secret scanning before commit
You decide and give the verdict. Each layer below feeds the agent or checks its work.

Who controls what

  • What to do

    The agent decidesThe prompt, AGENTS.md, skills it loads itself

    The harness enforcesHooks; skills only you can start

  • What it can touch

    The agent decidesA classifier that approves commands

    The harness enforcesDeny rules, the sandbox in strict mode, org-managed settings

  • When it's done

    The agent decidesEvidence you ask for, a goal check, a reviewer agent

    The harness enforcesTests, typecheck, a stop hook, CI

  • What it spends

    The agent decidesIts own pace, within the effort you set

    The harness enforcesA turn or spend cap

  • Who has the last word

    The agent decidesIts summary of what it did

    The harness enforcesBranch protection, a person who approves the merge

Let the agent decide most things. Enforce only the rules with no exceptions, and keep those few.

One change

  1. Research a read-only pass notes how the code works now, with file references
  2. Plan for an unclear goal, ask the agent to interview you; then steps, files and a check each, which you read
  3. Build one reviewable slice at a time; after two failed corrections, clear the context and prompt again with what you learned
  4. Verify tests and lint pass, the agent runs the app, then a fresh-context agent reviews the diff, against the plan too
  5. Review you read the tests first, then the riskiest part of the diff
  6. Merge a small pull request; static analysis runs in CI

a repeated mistake becomes a line in AGENTS.md; a rule with no exceptions becomes a lint rule or hook

Each step leaves a file or a check, so the next session starts from the repo. Clear the context before an unrelated change.

Verdicts

Adoptthe default now

  • AGENTS.md Holds the check commands and non-standard practices. Claude Code reads it when a repo has no CLAUDE.md.
  • A plan before code You read a written plan before the agent edits anything.
  • Auto permission mode In interactive sessions a classifier approves routine commands and can be wrong; "never" rules go in deny rules or hooks.
  • Checks decide "done" The agent runs tests, types, lint and the app before it reports back.
  • Fresh-context review A second agent reviews the diff before you do. Ask it to report correctness gaps and skip style.

Trialuse it on real work

  • Agent Skills Keep a few small skills and delete the ones you don't use.
  • Cloud agents Start a task from your phone or an issue; review the pull request it opens.

Assesstry it on the side

  • Agent-written workflows A script fans one task out to many sub-agents; check each one's output.

Cautionavoid it, or stop

  • Agent swarms Many agents on one goal. Costly and far from mature.
  • Repo overviews in AGENTS.md Leave them out; they didn't help agents solve more tasks.
  • Unattended long runs Give each one a check for done, a turn or spend cap, and a person who approves the result.
  • Maximum effort by default Max tends to overthink ordinary tasks. In Claude Code, medium for everyday work; high when easy-to-miss cases matter.
  • Prompt habits from older models IMPORTANT in capitals → plain words or a hook; "think hard" → the effort setting; "keep going" → a check to pass.
Rings from Thoughtworks' Technology Radar. The picks are this page's own.

Sources (12)