Recoleta Item Note
Graduate from Single-Session Coding: My Full Agentic Coding Workflow
This is an experience-based summary of an "agentic coding workflow" aimed at engineering practice. Its core goal is to upgrade single-session human-AI pair programming into a software delivery system that supports…
Summary
This is an experience-based summary of an "agentic coding workflow" aimed at engineering practice. Its core goal is to upgrade single-session human-AI pair programming into a software delivery system that supports parallelism, memory, and automation. The article emphasizes connecting planning, implementation, testing, review, and operations into a closed loop through worktrees, multi-agent orchestration, persistent task memory, and toolized prompts.
Problem
- Traditional single-session coding is single-threaded: when there is only one working copy, multiple agents will overwrite each other, making parallel development difficult.
- Long conversations consume context; after approaching or hitting the compaction boundary, agents become "dumber" and "lazier," and context rot / context poisoning can also occur.
- Many development activities still rely on manual handoffs: planning, implementation, testing, PRs, browser validation, production troubleshooting, secret management, and more lack a unified workflow, limiting efficiency.
Approach
- Use git worktrees + Conductor to give each parallel task an isolated workspace, allowing multiple agents to work safely in parallel; Conductor handles worktree creation, management, and cleanup, and supports switching between Claude/Codex within the same harness.
- Use Beads as an external persistent memory and task layer for agents: first write the plan as a markdown spec, then break it into beads (including dependencies). New sessions execute according to beads, a parent agent dispatches child agents in parallel, and each child agent commits code, closes its bead, and sends back notes after completing its task.
- Use Skills / AGENTS.md to encode high-frequency workflows and local conventions, such as brainstorm, beads-create, beads-implement, as well as project-level/directory-level development conventions and CLI tool usage instructions.
- Use a browser/CLI loop to extend the agent execution surface: use agent-browser or Browser Bridge for web/Electron validation, and directly handle PRs, production issues, deployments, and secrets through CLIs/MCP such as gh, Sentry, Railway, and Doppler.
- For model selection, adopt division of labor: the author prefers Codex for primary coding, while using Opus for code review, maintenance tasks, and local CLI chores.
Results
- The article does not provide rigorous experiments, benchmark data, or reproducible experiment tables, so there are no academically verifiable quantitative results.
- The clearest quantitative claim is that after using Blacksmith instead of GitHub Actions, the author says build times were reduced by about 50% (cut my build times in half), with a higher free tier as well.
- The author claims this workflow upgrades them from "pairing with one chat" to "running coordinated agents," enabling multiple agents to collaborate on planning, implementation, review, and maintenance, but provides no figures for throughput, defect rate, or cycle time.
- The author also claims to have shared this method with "several friends and peers" and repeatedly received feedback that "it works," but does not provide sample size, task types, or a comparison baseline.
- On model comparison, the article makes the strong claim that "Codex is clearly better than Claude at writing code," but does not provide benchmark names, scores, or task-level statistics.
- Overall, the article's main contribution is more about integrating a system-level workflow than introducing a new algorithm: it combines worktrees, task memory, child-agent parallelism, skill prompts, and browser/CLI automation into a "software delivery operating system."
Link
Built with Recoleta
Run your own research radar
Turn arXiv, Hacker News, OpenReview, Hugging Face Daily Papers, and RSS into local Markdown, Obsidian notes, Telegram digests, and a public site.