My thoughts on Personal Agent Protocol
1. I like that businesses can start with the website or APIs they already have. They don’t need to build their own agent to participate. Makes adoption a lot more realistic.
2. The permissions matter. Let my agent see my bookings without letting it cancel them. Give it enough access to do the job, and let me take that access back.
3. I’m pleased to see approval covers specific terms. If the terms change, the approval has to change too.
4. The business still takes the agent’s word about who approved. The spec says this directly. A business can have me confirm on its own site or app instead. Signed approvals are still on the roadmap but for trust this is necessary as I anticipate chargebacks and fraud to come heavy..
5. I’d like more options here. Finding a real person behind an agent is useful. Knowing that person approved this particular action is another question. We shouldn’t treat one check as answering both.
6. Same with per-person limits. A store might need to know I’ve already claimed my share, even if I switch agents. That shouldn’t require a shared identifier that lets every business follow me around the internet.
I want my agent to do more without supervising every click. However, businesses need enough confidence to let that happen. Poppy feels like a useful foundation to build on.
Today we’re releasing Personal Agent Protocol (Poppy) v1.0.
It’s on for everyone who uses Muse, @Meta's AI agent, enabled for every merchant on @Shopify, and backed by more than 100 design partners.
@zhangliang0115 perhaps include the commit SHA and have it explicitly call out anything it couldn’t verify, so the evidence stays tied to the actual change.
I pulled together lessons from the Claude Code team and our workflows at VerifyYou: from pre-push checks and CI to preview environments, agent review, and production monitoring.
I included prompts you can use to assess your own verification loop too https://t.co/EKucJV360F
@SydlaAI the builder can certainly propose changes when it discovers something and there should always be a retro against those discoveries to determine if it should be implemented after evaluating if it would product more positive outcomes for the overall system
@temitope_rotimi idk about that. x AI taking world by storm with grok bot imo Muse is also insanely popular based on download reports so worth putting Meta in the mix when you consider everyone else
@gagu_sh Thanks! imo that applies beyond Java too
make the checks you always want to run deterministic and easy to invoke with one command, then run the same checks in CI
the attention bottleneck really resonates... its a lot on my poor soul.
with that said, I want to measure unnecessary interruptions alongside important things the agent failed to surface.
fewer check-ins is only progress if it��s getting better at knowing when you actually need to be involved.
im depating making a game based on my families business (a tree nursery) I've already been building software for them to run their nursery operations at https://t.co/7PNsqAfA64 so am toying with the idea of making a game for it like farming simulator but for nurseries haha not sure where i would start though.. any ideas?
The lossy handoffs point really resonates. If two agents are constantly passing context back and forth, keeping that work in one loop with the right skills and tools feels like a better starting point.
I’d still keep independent verification separate where the stakes warrant it. And I’d want to compare more than the average score: which tasks broke, what did it cost, and did we preserve the same safety boundaries?
I’ve been digging into evals while thinking about how to build a better agent harness.
One thing I keep coming back to is that a successful outcome isn’t enough. We also need to understand what the agent did to get there, what changed, and whether we can trust the checks judging it.
Just published my top eval papers of 2026.
The list covers different aspects of evals to focus on: durability, quality, robustness, efficiency, etc.
It also covers new topics like verifiers, agent-as-a-judge, jev-as-a-judge, audit self-improvement, and more.