Just talk to it: a terse playbook for agentic engineering
Peter Steinberger published an update on his current Agentic Engineering workflow. Agent-driven development has moved from experiment to everyday workflow. In a practical account from an engineer working on a ~300k LOC TypeScript/React codebase (plus a Chrome extension, CLI, Tauri desktop client and Expo mobile app), agentic tooling handles most of the daily coding work. The central mantra: prefer direct, pragmatic interaction with agents rather than elaborate harnesses or choreography.
Core workflow highlights
- Primary model: gpt-5-codex on mid settings as the daily driver—chosen for a balance of speed and reasoning.
- Interface: a local codex CLI running several parallel agent sessions in a terminal grid, with agents making git atomic commits so commit history stays clean.
- Blast radius as a guiding principle: estimate how many files a change will touch and favor many small, isolated edits over large, sprawling changes to keep rollbacks simple.
- A lightweight approach to planning: ask for options or brief plans, stop agents mid-run if the observed work diverges, and iterate interactively.
Why codex over other agents
A concise comparison of the practical advantages observed:
- Larger usable context (~230k tokens) versus alternatives, which helps when the model must read many repository files.
- More efficient token usage, reducing context churn.
- Message queuing that enables batching tasks without having to craft elaborate steering prompts.
- Improved speed and stability in the CLI environment, contributing to a more productive loop.
- Better in-repo reading behavior—codex tends to scan more files before acting, which leads to more conservative and accurate edits.
Tooling and constraints
- Many auxiliary ideas—RAG, subagents, external harnesses, MCPs—are treated skeptically unless they add clear value; lightweight CLIs (including GitHub’s
gh) and standard tools liketmuxare often sufficient. - Agents are instructed via a large local Agents.md that encodes repo-specific preferences, coding patterns and git rules; this grows over time as the agent refines it.
- Practical habits: use screenshots in prompts for quick context, request tests after features, preserve intent with code comments, and run periodic refactor passes (agents do the heavy lifting).
The write-up is practical and prescriptive rather than theoretical: it focuses on what was changed in a working development loop, how agents are controlled, and which friction points matter in daily engineering. For the full workflow, examples and links to the author’s agent file and tooling notes, read the original post.
Read the full article: Just Talk To It - the no-bs Way of Agentic Engineering
