Anthropic has handed its Agent Skills standard to a neutral foundation. That's the one item today that should change a decision. The rest is agent plumbing: credentials, memory costs and runtime design. Worth knowing, but nothing to act on this week.
1. Anthropic gives Agent Skills to the Linux Foundation
What happened: Anthropic donated the Agent Skills standard to the Agentic AI Foundation (AAIF), part of the Linux Foundation, which it co-founded with Block and others.
Why it matters: If your teams have held off packaging internal know-how as Skills because it felt like a bet on one vendor, that reason just went away. A standard under Linux Foundation governance is much harder for any single lab to pull or bend toward its own models.
My position: start writing Skills now for your highest-repeat workflows, such as compliance checks, report formats and house style. Most of the work is documenting how your people already do things, and it carries over to any agent platform that adopts the standard.
Watch for fragmentation. A neutral standard only pays off if the other big providers ship support for it. Ask your vendors directly whether they will, and get the answer in writing.
2. 1Password lets agents log in without seeing your password
What happened: 1Password launched a feature that lets AI agents sign into a customer's personal accounts while keeping the underlying credentials hidden from the agent.
Why it matters: Agents need credentials to do anything useful, and today that often means pasting them into a prompt or a config file, which is exactly how they leak. 1Password is aimed at personal accounts for now, so don't treat it as your enterprise answer. The pattern is right, though. I'd ask whoever owns identity whether your stack can broker agent access the same way before agents touch production systems.
3. Memory is becoming the real cost of long-running agents
What happened: A new paper argues that as agent trajectories grow, key-value cache storage and I/O become the dominant cost, and runs can outgrow the model's context window entirely.
Why it matters: Long agent runs get expensive in ways token pricing won't show you.
4. Forward this talk to whoever runs your agent platform
What happened: At InfoQ, Manju Rajashekhar set out how the actor model covers state ownership, message passing and supervision for production agents.
Why it matters: Agents in production behave like distributed systems and fail like them. It's the right primer for anyone moving pilots into live use.
The bottom line: Today's one action is cheap. Pick two or three workflows your teams repeat every week and have someone write them up as Agent Skills. With the standard now under the Linux Foundation, that work is far less likely to be stranded by a vendor switch. Then put two questions to your platform and security leads. How do our agents authenticate without seeing raw credentials? And who is tracking what long agent runs cost us beyond tokens? If neither has a clear answer, that gap matters more than any model launch this month.
From the field
A practitioner note for anyone who owns the roadmap and the risk.
Most Multi-Agent Systems Should Be One Agent and a State Machine
Most multi-agent architectures we're asked to build should be one scoped agent plus deterministic orchestration. When handoffs earn their keep, and a test to decide.
— Daniel · usqrd.com · reply to this email, I read everything