Just Talk To It – the no-bs Way of Agentic Engineering | Peter Steinberger

Real Work with Agents — Practical Workflow Notes from an Agentic Engineer
This post summarizes a pragmatic workflow for agentic engineering drawn from a developer working on a large, multi-platform codebase.
It focuses on how to run parallel agents, manage commit blast radius, and prefer simple, reliable tooling over complex orchestration.
Why this matters to you.
• It shows how to get consistent results from agents by treating them as careful collaborators.
• It highlights operational practices that reduce risk when agents edit code.
• It challenges the impulse to over-engineer agent orchestration in favor of direct, observable control.

Header 1 — Agent-first day-to-day: terminals, parallelism, and atomic commits
The author runs multiple agent sessions in parallel inside terminal panes.
This keeps context visible and control immediate.
Why not a heavy orchestration layer?
Because seeing what each agent reads and writes simplifies debugging.
Example snippet to illustrate the flow.
Prompt to an agent: „Read repository files in this folder and propose a single atomic commit for feature X.“
Git behavior is designed so each agent commits exactly the files it changed.
That reduces merge noise and preserves a readable history.

Header 2 — Blast radius thinking and interactive steering
The author thinks in terms of blast radius for every change.
Small, focused edits are preferable to sweeping rewrites.
How do you steer an agent mid-run?
Hit escape and ask for a status update.
Example interaction.
Command: „Status update — what files will this change touch and why?“
If the agent drifts, abort or ask for options before continuing.

Header 3 — Model choice, context tricks, and tooling realism
The author prefers a pragmatic model that reads the repo and just gets to work.
Screenshots are used as quick context markers.
Many third-party tools come and go.
Why?
They often hide the terminal and obscure what the agent sees.
Example prompt pattern.
„Here is a screenshot of the failing test. Show me the likely cause and a focused patch.“
A simple, observable workflow wins over clever orchestration.

Wrapping up
Agentic engineering is no longer experimental for serious work.
But success depends on process, not just model novelty.
Keep changes small and visible.
Keep agents accountable with atomic commits and status checks.
For leaders, that means governance and measurable risk controls when you let agents modify code.
If you want a hands-on setup that prioritizes speed and clarity, consider a terminal-driven, parallel-agent workflow as a practical baseline.