Don’t let headlines fool you. AI isn’t replacing engineers... it’s helping them build a new class of products. One example I’m excited about is All Hands, an open-source coding agent that can take a scoped task and do real work toward it with minimal intervention.

What All Hands is

Think of All Hands as a tireless pair-programmer that can read a repo... plan a path... run commands... edit files... execute tests... and propose changes for you to review. You give it a clear goal and access to a sandboxed environment. It breaks the work into steps, executes them, and shows its receipts along the way.

You still own the system design and the merge button. The agent handles the grindy bits that slow you down.

Why this matters

• Speed on routine engineering... scaffolding, dependency bumps, config tweaks, boilerplate tests.

• Focus for humans... engineers spend more time on architecture, tradeoffs, and code review.

• Traceability... every command, diff, and test run is logged so you can audit the work.

• Repeatability... once a task pattern is dialed in, the agent can run it again and again.

What it’s good at

• Adding a feature flag and wiring it through config

• Migrating from Library A to Library B across a codebase

• Writing missing unit tests or snapshot tests based on examples

• Fixing linter and type errors introduced by a refactor

• Generating docs from code comments and updating a README

• Spinning up a small demo service to exercise an API

How a typical run looks

1. Scope the task
   Provide a concrete goal, constraints, and success checks.

   Goal: “Add Dark Mode toggle to the settings page... persist in local storage... include unit tests... pass npm test.”

2. Plan
   The agent outlines steps... files to touch... commands to run... and how it will verify success.

3. Execute
   It runs shell commands, edits files, and surfaces diffs. You can pause or nudge with a hint at any time.

4. Verify
   It runs tests and linting... and if something fails, it iterates until the checks pass or it reaches a sensible stop.

5. Propose
   It opens a PR or creates a patch with a summary of what changed, why, and how to reproduce the verification.

A simple task template you can reuse

You are a coding agent working in a sandboxed repo.

Goal:
- <clear outcome>

Guardrails that keep this sane

• Work in a branch... never on main.

• Limit scope... one feature or one migration at a time.

• Pin a test suite... the agent must make it green.

• Human review... treat outputs like a junior dev’s PR.

• Secrets discipline... provide only what’s required for the task.

Where to start

1. Pick one low-risk, high-repeatability task type.

2. Run All Hands locally in a sandbox on a small repo.

3. Capture your best task prompt as a template.

4. Track time saved... number of successful runs... PR rework rate.

5. Expand the surface area only after the metrics look good.

What this means for teams

Engineers aren’t going anywhere. They’re moving up the stack... spending more time on system design, reviews, and product outcomes while agents handle repetitive execution. Tools like All Hands make that shift real... today.

If you’re curious, start small... measure ruthlessly... and iterate on your task templates until the workflow feels boring in the best way.