“Anyone can borrow your tools. Nobody can borrow your taste.” — Nadina D. Lisbon

Hello Sip Savants! 👋🏾

Think about the last time someone on your team spent an afternoon teaching an agent how your company actually does things. On October 22, four companies that compete for many of the same customers each handed a piece of their agent stack to the same neutral home, the Linux Foundation’s Agentic AI Foundation [1], which now counts 305 member organizations less than a year after launch [2]. Then, on Monday night, the IETF closed submissions for its next meeting with a fresh stack of agent drafts in the pile [3][4][5]. The plumbing is becoming shared. The more interesting question is what stays yours. Three bites to start.

3 Tech Bites

🧳 Four rivals, one neutral home
- Anthropic contributed Agent Skills, an open format meant to make “specialized know-how portable across agent products and platforms,” and OpenAI contributed Open Responses, a shared request shape across several model providers based on its Responses API [1]. With Hugging Face’s OpenEnv and Alibaba’s OpenSandbox, the foundation now hosts 10 projects, up from three at its December 2025 launch [1].
🏛️ The IETF agrees to start
- The IETF approved its Agent Communication Protocols working group (agentproto) on October 8, with the goal to “define a common baseline agentic dialog management protocol” [3]. Thirteen drafts and revisions were posted between October 19 and the November 2 cutoff for IETF 127, which meets in San Francisco from November 14 [4][5]. None has been adopted yet, so think of it as an agreement to begin, not an answer.
📏 A measuring stick, still being calibrated
- A single-author preprint, not yet peer reviewed, audited the scoring machinery of one agent benchmark. After a scoring defect was corrected across 4,320 episodes, measured success fell from 94.3% to 65.5% [6]. The author keeps the conclusions narrow, but it’s a gentle reminder that measuring is still catching up.

Subscribe now

5-Minute Strategy

🧠 Your Two-Sentence Definition of Good

When the formats become shared, the judgment about what counts as good work stays with you. One option is to spend five minutes, on your own and before anyone hands you a dashboard, putting that judgment on paper:

Pick one agent-assisted workflow your team runs or is piloting.
Write one sentence describing what a good outcome looks like to you, in your customer’s terms rather than the system’s.
Write a second sentence naming one way a run could be marked “successful” and still disappoint you.
Open the score your team reports for that workflow and ask yourself whether it would catch your second sentence. (In the preprint above, one scoring defect was the difference between 94.3% and 65.5% [6].)
Bring both sentences to your next review as a question. “What would this score miss?”

The formats are becoming shared. Your definition of good is still yours to write.

Share

1 Big Idea

💡What Travels, What Stays

Every agent that works well inside an organization carries a little of that organization’s judgment. Someone sat down and wrote out how a refund should be handled, which system holds the real customer record, what tone works with a nervous client. A skill file is, in a very human sense, a team’s know-how written down. Until recently, most of that writing lived wherever one vendor’s tool happened to keep it.

That is the part that started to travel this month. Agent Skills was contributed specifically to make know-how portable across agent products, and Open Responses offers one request shape across several model providers [1]. Alibaba’s Xiaobin Xu put the spirit plainly: “we don’t want its future to be defined by a single company” [1]. For the people doing the teaching, it means their effort is more likely to outlast any single vendor decision. That’s a quiet form of respect for their work, even if no press release framed it that way.

The standards side is moving at a steadier pace, and that patience is worth appreciating. The IETF group plans to adopt a use-cases document by December 2026, an architecture by February 2027, and a protocol draft by March 2027, and its charter reuses existing identity and encryption building blocks rather than inventing new ones [3]. Shared foundations tend to take shape through careful, open conversation, and this one is still very much in the conversation stage.

What doesn’t travel is the judgment about whether the work was good. A standards body can agree on how a request is shaped or how a sandbox runs. It can’t decide what “good” means for your customers, your appetite for risk, or your team’s sense of craft. The benchmark preprint is a small, useful illustration: measuring turns out to be harder than plumbing, and a single scoring choice can change the story a number tells [6]. That isn’t a reason for worry. It’s a reason to keep a person close to the definition.

So the week suggests a pleasant division of labor. The formats can become common, and that’s good news for everyone who builds on them. The attention might go to the part that stays: the people who know how the work should be done, and the sentences that describe it. As everything around it becomes shared, that’s what will keep your organization recognizably yours.

What’s one thing your team checks for in agent work that no benchmark would think to measure? I read every response.

Leave a comment

P.S. If you know a leader whose team has been quietly writing down how the work really gets done, share this newsletter and help brew up stronger customer relationships.

Share TechSips with Nadina: The Newsletter

P.P.S. If you found these AI insights valuable, a contribution to the Brew Pot helps keep the future of work brewing.

Add to the Brew Pot! ✨

Resources

[1] Agentic AI Foundation Announces Four New Projects: Agent Skills, Open Responses, OpenEnv, and OpenSandbox

[2] Agentic AI Foundation Surpasses 300 Members in Less Than One Year

[3] Charter for Working Group: Agent Communication Protocols (agentproto)

[4] Agent Communication Protocols (agentproto): Documents

[5] IETF Important Dates

[6] When Agent Evaluation Becomes Unreliable: An Empirical Study of Oracle Failures and Configuration-Dependent Measurement in an LLM Agent Benchmark

Sip smarter, every Tuesday. (Refills are always free!)

Cheers,

Nadina

Host of TechSips with Nadina | Chief Strategy Architect ☕️🍵