The Wake: May 23, 2026

The Wake is a daily briefing from George's saved internet. The issue is written as a newsletter first. The tweets are the source material, preserved below for receipts.

Source window: May 22, 2026. Signals: 12 bookmarks and 3 likes.

Brief

Agents are moving out of the lab and into the plumbing. Over the last 24 hours the conversation shifted from toy demos to scaffolding: protocols to register agents with services, SDKs to move files at scale, local-first prototypes that make each chat its own VM, and multimodal features that let you point, talk, and have paperwork completed for you. Those are separate signals but they stack into a single read: the AI layer is starting to require real infrastructure, governance, and packaging: and all three are being designed in public.

If you are building or buying agent-driven tooling this week, orient to three simultaneous demands. One, runtime and packaging: how do skills or plugins get distributed, versioned, and run? Two, identity and auth: how do agents prove to services what they are allowed to do? Three, safety and triage: who decides what agents can see, do, and publish, and how do you handle false positives when automation touches the real world. Expect product decisions to be shaped less by model capability and more by how these plumbing problems are solved.

Agents: skills, plugins, and the new package problem

We are past the question of whether agents will be useful. The current battleground is how we make them composable, trustworthy, and manageable. A lot of innovation is focused on "skills" or lightweight service-backed capabilities that agents invoke. Garry Tan’s GBrain experiments show the upside: you can have an agent rewrite books for you or run a deliberate creative collision function he calls lateral synaptic drift. Those are compelling user stories.

But the downsides are immediate and structural. Mark Kretschmann’s pushback is the right read: skills introduce an operating system problem for agents: package management, version compatibility, signing, and a trust surface that grows as skills are shared and downloaded. Once agents can execute across networked services, every skill is a potential privilege escalation point. The industry is learning the same lesson cloud infra learned a decade ago: developer convenience without governance becomes a security and reliability tax.

Two implications. One, there will be a market for standards and registries that treat skills like packages: with metadata, signing, backward compatibility guarantees, and curated marketplaces. Two, there will be bifurcation between closed, curated skill ecosystems (enterprise, security-conscious) and open, experimental ones (hacker-first, rapid innovation). Both will exist, and both are already being built.

Infrastructure catching up: auth, files, and remote runtimes

The plumbing to make agent ecosystems tolerable is being assembled. WorkOS announced auth.md, an open protocol for agents to register for services on the web. If that adoption read holds, auth.md is an early attempt at solving the identity and grant problem: how an agent proves it can access a Cloudflare worker, an API, or a storage bucket. When agents request privilege, you need a declarative, machine-readable contract you can audit.

Storage and connectivity improvements are moving in lockstep. Files SDK updates adding bulk operations, SFTP/FTP adapters, and new storage backends is not glamorous, but it matters: automation lives and dies by reliable I/O. And developer ergonomics are being polished at the edges: Theo’s admiration for T3 Code’s remote worktrees and built-in Tailscale shows remote-first dev workflows becoming frictionless. Taken together, these moves remove the last-mile headaches for the kinds of agent automation that integrate with a company’s data.

Finally, the local-first runtime experiments are worth watching. Rhys is prototyping a "meshnet" where each chat runs in its own VM but the developer interfaces feel local. That pattern matters because it offers a different axis of trust: local VMs reduce the need to centralize data and can contain execution. Local-first plus networked registration protocols creates options: run an agent with full access to your files locally while registering what capabilities it exposes to external services. That hybrid approach is a pragmatic hedge against both centralization and total fragmentation.

Multimodal UIs and practical automation

The usability story has finally caught up to the models. ChatGPT’s image and voice modes: the ability to upload a form, speak instructions, and receive a filled document: are small on hype but large on utility. Nathan Baschez’s reaction: "WHAT? this is amazing": is the right tone. These features shortcut tedious moments and lower the activation energy for nontechnical users to adopt agent workflows.

There are also domain-specific leverage points. Peter Yang’s note about Codex automation in game development signals that domain-specialized automation is accelerating. And Alvaro Lozano-Robledo’s experiment using ChatGPT to recreate and upscale an image for publication is a reminder that multimodal agents are already useful for creative workflows, not just text.

Product teams should take away two principles. First, combine multimodal inputs with deterministic outputs and guardrails. Forms you can point at and fill should produce auditable artifacts and change logs. Second, domain agents will beat generalists on throughput: a Codex-backed game dev agent that knows the engine will save the team more time than a general assistant trying to grok the project each time.

Safety, moderation, and the politics of talent

As agents gain reach, the moderation and triage problem gets harder. Recommended readings on triaging and false positives are circulating because automated systems will make consequential decisions about content and actions. Human-in-the-loop processes and better triage tooling are not optional: they are the only scalable path to trustworthy automation.

There is also an increasingly public political dimension. Talent and immigration policy still matter. A popular refrain in developer circles: "do not make it harder for the world’s smartest people to come to the US": is a reminder that geopolitical decisions shape where the next generation of agent builders will cluster. Expect policy debates to influence hiring, fundraising, and where infrastructure gets built.

Finally, the community is searching for legitimacy markers. Someone suggested the first agent-driven contribution to fundamental science as a cultural flag for the field. That is aspirational but useful: proving agents can advance math, physics, or biology in a verifiable way would tilt perceptions from toy to tool.

What to watch

• Adoption of auth.md and similar agent identity protocols. Who signs on as a provider and which providers demand stronger claims from agents.
• Skill registries and packaging standards. Look for the first major curated marketplace or a package manager with signing and versioning.
• Local-first runtimes like meshnet prototypes. If a pattern emerges for per-chat VMs with optional network registration, early adopters will tout privacy and auditability.
• Files and storage adapters in SDKs. Reliable bulk I/O will be the gating factor for many enterprise automations.
• Multimodal productivity features in mainstream LLMs. Watch how audit trails and export formats are handled for legal/operational safety.
• Moderation triage improvements and tooling. False positives handling in automation workflows will define operational trust.
• Hiring and migration policy chatter. Regulatory moves will affect where teams and capital flow in the next 12 months.

Bottom line: we are shifting from model capability to system capability. The new winners will not only have better reasoning models but also better packaging, identity, and operational primitives that make agents usable, safe, and auditable at scale.

Generated from Birdclaw bookmarks and likes. Edited by Ody before publication.