Skip to content

SYSTEM Cited by 2 sources

GitHub Copilot

Definition

GitHub Copilot is GitHub's AI coding assistant, originally an IDE autocomplete tool and later an agentic offering. It is a subscription-based product distinct from raw model API access. (Source: sources/2026-08-13-zalando-agentic-engineering-at-zalando-a-snapshot)

One shared harness across surfaces

Multiple Copilot products — Copilot CLI, the GitHub Copilot app, and Copilot code review — run on the same underlying agent harness, so harness-level efficiency changes propagate across all of them. GitHub still re-measures each change per surface, because a change that helps one workflow can raise cost in another (workload-local-evidence). (Source: sources/2026-09-02-github-how-we-make-ai-coding-more-cost-efficient)

Cost-efficiency engineering (2026-09-02)

GitHub shipped four independent harness changes that cut per-task AI cost with no detected quality regression, unified by one thesis: optimize the completed task, not the individual tool call (the local metric trap). Each was validated with task-level measurement — offline agentic-coding benchmarks, then online A/B experiments.

Change Mechanism Effect (AI-credit metric)
Selective output compaction Classify output; preserve source-like/arbitrary, reorganize search losslessly, compress build/test/lint noise; keep a recovery path ~5.5%
Remove view line-number prefixes Drop unused per-line numbering (current editors match code, not numbers) ~3.1% (~5% offline inference)
Compact the task-tool prompt Self-rewrite prompt ~50% smaller, guarded by a behavioral regression test ~2.9% (~1,300 tokens/turn)
Reduce notification round-trips Batch background shell/sub-agent completions and deliver results directly ~2.3%

The task tool launches specialized agents for parallel work; compressing its prompt once serialized independent agents until a one-sentence rewrite ("Independent agents can run in parallel; consider side effects") restored parallelism. GitHub also evaluated the external RTK shell-output shortener and found it raised end-to-end cost in their harness via reread/rerun recovery.

Seen in

  • Zalando — an early Copilot user "from the early days when it offered autocomplete in the IDE." To complement it with API-based model access, Zalando built its LLM proxy in Jan 2024. Tools like opencode and pi are valued for mixing a Copilot subscription with the API proxy in one workflow.
Last updated · 766 distilled / 2,225 read