Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

1 Commit
 
 
 
 
 
 
 
 

Repository files navigation

Working with a Coding Agent — a study and a playbook

A close reading of thirteen months of real, day-to-day coding-agent sessions, turned into practical guidance for both the human and the agent.

What's here

File What it is Audience
index.html The write-up / blog post. The story, the data, the six-move loop, the autonomy question, with diagrams. Start here. Anyone
human-playbook.html A one-page field guide: the loop, three checklists, a phrase bank, an "earn more autonomy" section, and anti-patterns. Built to print. The human
agent-instructions.md A tool-agnostic instruction set to drop into any agent's standing-instructions file (AGENTS.md, a rules file, a system prompt). The agent

Open the two .html files in any browser. They're self-contained (no internet, no assets, no tracking) and safe to share.

What it's based on

  • 211 locally-stored agent sessions (154 with real back-and-forth), 1,436 human turns in total.
  • Spanning April 2025 – June 2026 (~13 months) across 19 codebases — a mobile + embedded-device app and SDK, a regulated reporting tool, a data/metrics pipeline, a server framework, even coursework and sessions reading the agent's own source.
  • Plus 17 deep 2026 "agent-mode" transcripts capturing ~3,400 tool executions, used for the tool-usage and autonomy-depth figures.

How it was done

  1. Combined two storage sources so the timeline is complete: exported session logs (.jsonl, the 2026 sessions) and the live workspace databases (.json sessions, reaching back into 2025). They were deduplicated by session id — there was no overlap, so the union is the full record.
  2. Parsed every session, reconstructing each conversation's turns (handling both whole-session snapshots and the streamed, incremental save format).
  3. Profiled the corpus quantitatively — models, modes, turn counts, tool usage, the 2025→2026 shift in agent depth, and the share of turns matching a dozen behavioral patterns.
  4. Read the conversations — the deepest sessions in full, plus a categorized harvest of exemplar quotes across all turns, spanning every domain and both years.
  5. Synthesized the recurring behaviors into a six-move loop, a human playbook, a mirror set of agent principles, and an evidence-based answer to the autonomy question.
  6. Verified every load-bearing quote against the source transcripts and scrubbed all product names, identifiers, paths, and people for public sharing.

Two findings worth pulling out

  • The stance is constant; the leash grew. Agent depth rose from ~3.3 to ~9.5 tool-actions per human turn between 2025 and 2026, yet the human's grounding-and-verifying style is identical in the first session and the last.
  • Autonomy is earned through verifiability. The way to let the agent do more on its own isn't to instruct it to be bolder — it's to make more of the work cheap to check (tests, golden files, green builds, standing rules) and keep a hard stop at anything irreversible.

The one-line takeaway

Treat the agent as a fast, capable partner that has no idea what actually matters. Give it the missing context generously; verify its claims relentlessly. Hand it the fast part of the loop — keep the judgment for yourself, and make the judgment cheaper to check over time so the fast part can grow.

All examples are anonymized; quotes are lightly edited for typos only. Independent of any other analysis in adjacent folders.

About

Working with a Coding Agent — a study and a playbook

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages