Loops, Harnesses, Squads, Hill Climbing: 2026's AI Jargon Translated Into Plain English

The last six months produced a pile of new AI words: loop, harness, squad, fleet, hill climbing. They sound technical and describe fairly ordinary things. GitHub published an official glossary on 2026-09-02; this piece translates it and adds the “so what do I do” for each. Verified 2026-09-08.

1. Four words about directing agents

Loop engineering

Official definition: “the practice of designing repeatable systems around agents, instead of manually prompting them for one task at a time.”

Plainly: moving from “every morning I ask the AI to review my issues” to “that happens every morning by itself.” GitHub’s own example is automated daily issue reviews on a schedule.

What to do: find something you prompt an AI for three or more times a week and turn it into a schedule or script. Lowest effort, most direct payoff on this list.

Ralph loops

Official definition: one implementation of the loop concept — “you give an agent a detailed task, often from a product requirements document or spec, and have it keep working until the job is done.”

Plainly: not a conversation, a delivery. You hand over a clear spec and it loops until finished.

The catch: success here depends on how clearly the spec is written, not on how clever the model is.

Squads

Official definition: “a group of agents with different roles,” usually mirroring a real team — planning, vetting, implementation, testing, review.

Plainly: instead of one agent playing every part, several agents each hold one.

Fleets

Official definition: “parallel agents working on tasks at the same time,” able to run alongside squads either sequentially or in parallel.

Plainly: a squad divides labour; a fleet runs many at once. The two stack.

2. Two words about the equipment around the agent

Harness

Official definition: everything surrounding the model “that makes it useful in your workflows” — tools, permissions, memory, context, orchestration.

Plainly: the model is the engine; the harness is the rest of the car. People reach for a stronger model when the harness is what is missing.

What to do: next time an agent feels useless, ask these three before switching models:

1. Does it have the tools it needs (run commands, read files)?
2. Does it have the context it needs (spec, existing code, conventions)?
3. Does it remember what was decided last time?

Hill climbing

Official definition: “the process of improving agents and harnesses over time” through evaluation feedback and iterative tooling adjustments.

Plainly: not tuning to perfection once — measure, change a little, measure again. The name is the metaphor.

The load-bearing word is “measure.” Without evaluation, each change just swaps one feeling for another.

Forward deployed engineer

Official definition: a customer-facing software engineer, sales engineer or solutions engineer, “often with an AI focus”, helping teams integrate AI tools into existing systems.

Plainly: a new job title — and its existence says something. The hard part of AI tooling has moved from “the model is not good enough” to “it does not fit our systems.”

3. Three kinds of “open”, and they are not interchangeable

The most misused trio, defined cleanly:

The middle one causes the confusion. Many models described as “open source” are only open weight — you can run them, but you cannot see what they were trained on. For compliance and auditability, that distinction is the whole point.

4. How to actually use this glossary

Do not memorise it. When someone uses these words, use them to ask yourself three better questions:

1. Am I doing this by hand every time? (→ make it a loop)
2. Am I complaining about the model, or about the harness? (→ usually the harness)
3. How would I know if my change helped? (→ no evaluation, no hill climbing)

Those three questions beat memorising seven pieces of vocabulary.

Related: eight steps for evaluating an LLM before production, money guardrails for AI agents, and the GitHub Copilot tool page.

Source read directly on 2026-09-08: GitHub’s official post Decoding the new AI lingo: Loops, harnesses, squads, hill climbing… oh my! (2026-09-02). All ten definitions are quoted from it. The “what to do” notes and the three questions in section 4 are ours, not GitHub guidance.