Principal, Member of Technical Staff at @Microsoft | @GitHub. Building AI-native developer experiences across GitHub Copilot Agentic Tools. Husband, father.
The best result doesn’t have to cost the most.
GitHub Copilot Fusion chooses and coordinates models for each task, so developers on GitHub can focus on the work.
On Terminal-Bench 4.1, Copilot Fusion Max achieved a 71.2% resolution rate, 5.4 points ahead of leading frontier models Opus 5.5 and Fable 5.2, while costing 88% less than Fable 5.2 and 57% less than Opus 5.5.
If an agent keeps making the same mistake, don’t just add another paragraph to the prompt. Turn the invariant into an API constraint or an automated check. What’s one rule you can enforce today?
.@poteto's talk at AIE Code is excellent.
If I had to pick one still from that felt genuinely original, it would be this one:
"nuance tends to lead to slop"
To increase quality, design API's and automated checks that remove the nuance from agents' decisions.
- Ban useEffect
- Ban 99% of explanatory comments (they just lead to workarounds)
- Create API's that hide complexity from agents
Full talk below.
An agent task board isn’t an operating model. Decide which work needs a tight feedback loop, which can run unattended, and where a human must approve. What makes you switch modes?
The future of software engineering is not task boards. It’s a mix of being able to organize work “in the small” and “in the big.” There will be times when you want to work on something deeply with an agent, and there will be times when you want to be completely hands-off. This will vary from team to team, and new ways of working will grow out of it.
Code scanning alerts, right inside the Copilot app. GitHub’s security extension is in public preview. Pick one alert, ask for an explanation, review the fix.
Start here: docs.github.com/en/copilot/h…
A red CI job should still leave evidence. Waza v0.41.0 includes the fix for JUnit output after completed grading failures; the run stays nonzero. Test that contract in your own suite.
microsoft.github.io/waza/rel…
Before trusting an eval gate, make it fail on purpose. Waza v0.41.0's offline Quick Start shows a passing text check and an impossible-predicate control. No model calls; no claim about agent quality.
microsoft.github.io/waza/qui…
If you attending #MSIgnite in San Francisco this coming week, I'll be giving a talk about Microsoft Autopilot!
Autopilot is our autonomous digital teammate that is powered by @openclaw and runs securely in the Microsoft 365 Cloud.
Hope to see you there!
Automate the Follow-Through with Autopilot ignite.microsoft.com/en-US/s…
We're hiring! Come help us define the AI Agent platform of the future. I've never worked at a more exciting time where the combination of technology, opportunity, and innovation is coming together 🥳
We're hiring for a variety of roles and levels around agents + data, community and outreach, developer experience, platform, and more.
I don't always get the chance to reply to all of my DMs but feel free to drop me a note with your information / experience as well - but best bet is to apply at some of our open roles:
apply.careers.microsoft.com/…apply.careers.microsoft.com/…
You can now sign-in ONCE to services like Figma, Sentry, Slack, or WorkIQ, and the connection is available immediately on every GitHub Copilot surface, on every device. No need to re-authenticate.
A special thanks to first-time contributors Ivan Dijan and Erik Galler for improving MCP mocks, grading evidence, responder reliability, and the Lab experience.
Microsoft Copilot and agents delivers better outcomes \ $ vs. ChatGPT and Claude. Why? 1. We optimize model usage. 2. We deliver AI native APIs and data enrichment.
Today's latest news is we are bringing the #SharePoint AI tools into Work IQ for a range of solutions.
techcommunity.microsoft.com/…
Ran an experiment that ran for a day that resulted in ~90 PRs to improve the performance of ASP.NET Core:
github.com/dotnet/aspnetcore…
Fun facts:
- These were using Github's new cloud agents support that shipped this week at GHU.
- Opus 5.5 is very good
- ~$230 in tokens.
#dotnet#aspnetcore
GitHub has released Copilot Fusion in GitHub Copilot, which orchestrates between Anthropic and OpenAI models. It scores 2 points below Opus 5.5 in Claude Code at half the Cost per Task in the Coding Agent Index
GitHub describes Copilot Fusion as a runtime orchestration system, which selects and coordinates models to optimize performance, cost, and latency for each task. Copilot Fusion chooses between a number of models from Anthropic and OpenAI, including Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Astra, and GPT-6 Sol. We benchmarked its max setting in GitHub Copilot CLI in the Artificial Analysis Coding Agent Index v2.0.
Token costs are charged based on the underlying models used to complete a request.
Key takeaways:
➤ Cost-efficient for near-frontier performance: Copilot Fusion in GitHub Copilot costs $5.59 per task, versus $11.42 for Claude Code with Opus 5.5 (max). Devin Fusion with Opus 5.5 (max) + GPT-6 Luna (high) scores 62 at $8.00 per task.
➤ Slight lead in Terminal-Bench 4.0: The Coding Agent Index v2.0 weights its three constituent benchmarks equally. Copilot Fusion in GitHub Copilot leads Terminal-Bench 4.0 with a score of 71. It ranks third on FrontierCode in the chart at 51, behind Claude Code with Opus 5.5 (max) at 56 and Devin Fusion with Opus 5.5 (max) + GPT-6 Luna (high) at 54. Its SWE-Atlas-QnA-AA score lags at 57, between GPT-6 Astra in Codex and Gemini 4 Argon in Antigravity CLI.
➤ Half the token use of Opus 5.5 in Claude Code: GitHub Copilot Fusion used 7.2M tokens per task in the Coding Agent Index, approximately half of the tokens used by Claude Opus 5.5 in Claude Code, and 30% less than Devin Fusion with Opus 5.5 (max) + GPT-6 Luna (high).
I could watch this all day! Agent Radar might be the best way to show all of my agents running! (yes its animated too!)
Open GitHub Copilot app and start a new chat - "install and open this canvas - github.com/github/awesome-co…"
Amazing work from @awlnx