$ today_in_10_seconds
GitHub published controlled benchmarks pitting its Copilot CLI harness against Claude Code and Codex CLI on identical models and tasks. The headline: comparable task-resolution rates, but lower token consumption across most configurations. The quiet argument underneath it is that the harness — not the model — is becoming the thing worth competing on.
$ lead_story
GitHub benchmarks its agentic harness against the model vendors' own
GitHub ran Copilot CLI head-to-head against the harnesses that ship these models natively — Claude Code for Sonnet 4.6 and Opus 4.7, Codex CLI for GPT-5.4 and 5.5 — holding model, task, context window, and reasoning effort fixed across SWE-bench, TerminalBench, and internal benchmarks. The result is effective parity on task resolution with lower token spend in most cases, which matters because it isolates the harness as an independent variable: the scaffolding around the model (tool orchestration, context handling, routing) measurably changes cost and outcome even when the intelligence is held constant. For anyone building agents, this is the clearest public signal yet that harness engineering is a real performance lever, not plumbing. And GitHub's pitch — 20+ models behind one harness, plus cross-model tricks like one model critiquing another — is a bet that vendor-agnostic orchestration beats being locked to a single lab.
// why_it_matters: When the model is a commodity you can swap, the harness is where the differentiation — and the token bill — actually lives.
$ also_worth_attention
Getting more from each token: context handling and model routing
GitHub's companion post on the optimizations behind the benchmark numbers — how Copilot trims context and routes between models so more of each session goes to useful work. The agentic takeaway: token efficiency is increasingly an orchestration problem, solved above the model layer, not inside it.
link →Take your local GitHub sessions anywhere
Remote control for Copilot sessions is now GA — start an agent run in VS Code or the CLI and hand it off to your phone. As agent tasks stretch into long-running, multi-step jobs, the interface is decoupling from the machine that kicked them off, which is what real delegation looks like.
link →Scanning for vulnerabilities with GitHub Security Lab's open-source AI framework
GitHub's open-source Taskflow Agent hunts auth bypasses, IDORs, and token leaks with reported high accuracy on high-impact classes. A concrete case of a narrow, tool-using agent doing focused security work rather than chasing general-purpose autonomy.
link →
$ bigger_picture
The trend behind the headlines
For two years the conversation was almost entirely about models — bigger context windows, higher benchmark scores, the next frontier release. GitHub's data quietly reframes that: with the model held fixed, the harness around it still swings cost and task completion enough to matter. That's a maturing of the field. The differentiation is moving up the stack into orchestration, routing, and context management — the unglamorous engineering that decides whether raw intelligence actually ships working code. If you're building agents, the lesson is to stop treating the harness as glue and start treating it as product.
GitHub Copilot SDK
The shared component powering Copilot CLI, the Copilot app, and Copilot code review, now the foundation for the harness GitHub just benchmarked. If you want to build your own surface on top of a multi-model agentic harness instead of writing orchestration from scratch, this is the layer to look at.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime