Uber Scales MCP, OpenAI Updates Agents, and Harness Scores Split
2026-10-04
Uber is tackling the messy part of enterprise agents: giving them access to thousands of internal services without losing control of discovery or security. Meanwhile, the same model reportedly scored 62% in one agent harness and 33% in another. That gap should make every leaderboard reader pause.
Today: Uber’s internal MCP infrastructure, the agent harness score gap, and AI infrastructure debt.
🧠 LAUNCH
Pi pod brings coding agents to self-hosted sandboxes. It runs Pi agents on a server you control, giving teams an option for isolating agent work. If your agents need their own execution environment, this is worth a trial. (70 points | 27 comments) Read more →
Kolibri-1 appears on Hugging Face as a text-generation model from Aleph Alpha. Its 243 likes show early interest; the supplied snapshot shows zero downloads, so inspect the model card before treating attention as adoption. (243 likes) Read more →
🔧 TOOL
OpenAI points builders to new Agents API capabilities
Your agent integration may have new options to evaluate. OpenAI is directing developers to an Agents API update, though the supplied post does not spell out the changes. Check the update against what you are building before changing your implementation. (442 likes | 24 RTs) Read more →
Claude Code mods make capability discovery personal. The “You should know” example helps users keep track of what Claude can do, showing a practical use for mods beyond customization for its own sake. Try it if useful features keep slipping past your team. (746 likes | 26 RTs) Read more →
Uber maps out MCP at enterprise scale
Connecting agents to internal tools gets harder when those tools span thousands of services and many teams. Uber’s MCP approach, described in this post, addresses discovery, security, and operational consistency across that sprawl. Those are the design decisions to study if agent access is growing faster than your tool inventory. (1,034 likes | 107 RTs) Read more →
Pi adds Nix installation and project-level MCP overrides in v1.0.1. The Nix flake supports macOS and Linux, while project overrides let developers vary MCP server access by project. That is a useful control when one global tool setup no longer fits. (158 likes | 5 RTs) Read more →
Claude Code updates shared agent coordination in version 2.1.289. Teammates can spawn shared agents through agent.spawn; the release also unifies agent IDs and clarifies idle and waiting states. Check the changelog if your workflow depends on multiple agents working together. (30 likes | 3 RTs) Read more →
📝 TECHNIQUE
Opus 5.5 gets a practical usage guide for Claude and Claude Code. The useful move is to test its advice on your own tasks, where you can see whether a change improves the work you actually ship. (146 points | 103 comments) Read more →
GPT-5.5 pins need an audit before October 14. A post flags that date as its retirement from ChatGPT and Codex; custom agents, scheduled tasks, and scripts with the model name hardcoded deserve a check now. (35 likes) Read more →
🔬 RESEARCH
One model, two harnesses, a 29-point score gap
An agent benchmark can change dramatically without changing the model. Hugging Face reports that identical model weights scored 62% in one harness and 33% in another. Its guide describes training across multiple harnesses through a proxy approach, a useful direction for anyone trying to build agents that work beyond one evaluation setup. (332 likes | 58 RTs) Read more →
The agent claimed success; the database told another story
A confident “done” is worthless if the underlying state says otherwise. Microsoft’s ThinkingBox post examines an agent completion claim that disagreed with database state. For builders, the lesson is concrete: verify the result in the system the agent was supposed to change. Read more →
Wavestone AI Lab proposes a seven-part blueprint for agent harnesses after reportedly comparing 11 agents, including Claude Code and Codex. The framework could give builders a sharper checklist for system design and evaluation. Read the paper before adopting its categories as settled fact. (125 likes | 8 RTs) Read more →
💡 INSIGHT
Airbnb applies AI across product development and the guest experience, according to an interview with Ahmad Al-Dahle. The interesting question is how those uses connect inside an established business, where changes must work for both internal teams and guests. Read more →
AI hiring claims need a narrower lens. The cited assessment says AI may be affecting hiring for exposed junior white-collar roles, while attribution remains contested and aggregate labor-market claims go further than the evidence. Treat a shift at the hiring margin as a serious signal, not proof of economy-wide displacement. (225 likes | 31 RTs) Read more →
AI data center debt is becoming a market story of its own. Credit analysts and asset managers are increasingly treating the financing as a distinct category, according to The Information. Watch it alongside model releases: access to capital can shape how quickly providers add compute. Read more →
🏗️ BUILD
Pi provides the harness in a guide to building Dots-style assistants, paired with a custom gateway. It is a useful architecture to inspect if you want to understand the pieces behind an assistant rather than only its interface. (235 likes | 27 RTs) Read more →
Claude Opus 5.5 powers a proposed motion design studio setup using Claude Code, Node, ffmpeg, Playwright, and project rules. The author’s key point is that the harness matters more than a clever prompt; try the steps on a small video project and judge the output. (47 likes | 4 RTs) Read more →
🎓 MODEL LITERACY
Agent harness effects in evaluation: A harness is the surrounding system that gives a model tools, instructions, and a way to act across steps. Change that system and the model can perform very differently: the same weights reportedly scored 62% in one harness and 33% in another. When someone cites an agent benchmark, ask how the agent was equipped and run. The score measures that setup as well as the model.
⚡ QUICK LINKS
- OpenAI plugin extensions: DevDay takeaways for developers building extensions. (151 likes | 15 RTs) Link
- OpenAI safety departure: A safety leader quits and publicly criticizes the company’s culture. (60 points | 14 comments) Link
- Sam Altman on AI judgment: A warning against giving models religious authority or surrendering human judgment. (20,569 likes | 1,371 RTs) Link
🎯 PICK OF THE DAY
The 29-point harness gap should change how we read agent scores. Hugging Face reports the same model reaching 62% in one agent harness and 33% in another. That is too large to tuck into a footnote about evaluation setup. A model’s agent score measures its harness as well as its weights, and training and evaluating across harnesses may tell us more than another isolated benchmark gain. Builders should demand the harness details before letting a leaderboard decide their stack. (332 likes | 58 RTs) Read more →
Until next time ✌️