Summary of an article by Cassidy Williams, GitHub Blog (September 2, 2026), based on a GitHub Podcast episode.
Loop Engineering
Building repeatable systems around AI agents instead of manually prompting one task at a time.
Example: a scheduled loop that fetches issues, has an agent summarize them, checks the output, and escalates problems — basically an "AI-native cron job."
Ralph Loops
A rougher version of loop engineering. You give an agent a detailed task (like a spec document) and let it keep working until done.
- Useful for breaking big tasks into repeated plan-act-check cycles
- Can burn a lot of tokens and compute, since every retry costs more
Squads and Fleets
- Squad: a group of agents with different jobs (one plans, one checks the plan, one builds it, one tests, one reviews) — like a real team
- Fleet: multiple agents working in parallel
You can run a squad within a fleet, either in parallel or in sequence. The point is splitting work by role instead of one agent doing everything.
Harnesses
Everything around the AI model that makes it actually useful: tools, permissions, memory, context, orchestration.
Named after horse harnesses — the model is the horse's raw power, the harness directs it safely. GitHub Copilot is an example: it connects the model to codebases, editors, pull requests, and terminals.
"Harness engineering" means building and improving that surrounding system.
Hill Climbing
The ongoing process of improving agents and harnesses using feedback and evals.
Example: checking whether an agent reviewing pull requests actually finds real bugs, then adjusting the tooling to get better results.
Forward Deployed Engineer
Not really a new role — it's a customer-facing engineer (sales/solutions engineer) who helps clients integrate technical solutions. Now often focused specifically on AI tools and agents.
Closed Models vs. Open Weights vs. Open Source Models
| Type | What's available |
|---|---|
| Closed models | Only API/product access; no weights, data, or training process (most big frontier models) |
| Open weight models | Trained weights are downloadable and can run locally; training data/method may stay private |
| Open source models | Model, code, data, and training process are all available for inspection and modification |
Takeaway
These terms will keep evolving — some will stick, some won't. What matters more than the buzzwords is the underlying practice: can your workflow repeat reliably, how do you validate output, how much human oversight is needed, and how do you keep improving the system.
Source: github.blog
Top comments (0)