loop engineering AI is the new dev stack

loop engineering AI just got a name from GitHub, and it describes work developers are already shipping.

StartupHub data

Companies working on this

Profiles of the companies named in this story, with founding year, headquarters, and a short description from our database.

Loop

A social platform designed for matchmaking and personal connections.

Founded

2015

Funding

$95M

The Sept. 2 post by Cassidy Williams, host of the GitHub Podcast with Marlene Mhangami and GPS, maps the jargon flooding PRs and standups. Some terms name real patterns. Others rebrand old jobs.

How loop engineering actually works

Loop engineering replaces one-shot prompts with repeatable systems. Think cron job with judgment.

Instead of asking an agent each morning to triage issues, you schedule a loop. It fetches issues, passes them to an agent, validates the output, and escalates failures. Ralph loops are the brute-force cousin, running plan-act-check cycles against a spec until it's done.

They work, but burn tokens, context, and compute fast. Well-designed loops add skills, observability, routing, and checkpoints so you're not stuck saying "try again."

A harness is everything outside the model that makes it useful: tools, permissions, memory, context, and orchestration. The post uses a horse analogy and points to GitHub Copilot as a harness that wires models to editors, codebases, terminals, and pull requests.

Squads and fleets describe multi-agent work. A squad splits roles like a real team, with a planner, implementer, tester, and reviewer. A fleet runs agents in parallel. You can run a squad as a fleet.

Hill climbing is the improvement loop. You eval whether the agent finds real bugs or writes useful fixes, then tune the harness and measure again.

Why harnesses expand the trust boundary, and what's still open

Every harness is a new trust boundary. Loops that can read issues, write code, and open PRs need permissions, and too many permissions turn prompt injection into code execution.

GitHub's guide flags validation and observability as primitives but offers no standard for either. Hill climbing for quality is not the same as hill climbing for safety.

Squads make this harder. Parallel agents share context and tools, so one compromised agent can poison others without strong isolation and checkpoints.

Startups are where this gets tested first. Anysphere with Cursor and Replit with Agent now compete on harness engineering, not model choice. Whoever ships least-privilege loops and auditable checkpoints wins enterprise deals.

The post also clarifies model openness. Closed models are API only. Open weights can be downloaded and run locally but hide data and training details. Only open source exposes weights, code, data, and process.

More openness helps you audit a harness, but it doesn't fix a loose one. And forward-deployed engineers, the rebranded sales engineer now embedding AI workflows with customers, are the human stopgap until the tooling matures.

Buzzwords will fade. Loops with proper permissions will not.