The Wake: May 17, 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 16, 2026. Signals: 10 bookmarks and 4 likes.

Brief

Codex in ChatGPT is no longer just an assistant, it is becoming the developer. Over the last 24 hours we saw a cluster of announcements and reactions that together read like a prototype for how software will be made, shipped, and secured in the era of agentic tooling: a model that can edit, test, and run code on-device; infrastructure to let agents act without ever holding raw credentials; new memory architectures for effectively infinite context; and people building languages and toolchains tuned for agents. The plumbing is brittle but the direction is clear: lower friction, tighter feedback loops, and a new set of security and labor tradeoffs.

Codex as a bootstrapper

What changed is not a single feature, it is a workflow. People are using Codex inside the ChatGPT iOS app to generate code, run snapshot tests, and iterate on the implementation with the same model doing the heavy lifting. Developers report being able to build and run projects on a paired iPhone without a laptop, and to ask the model for snapshot tests and immediate previews (@Dimillian). That matters because the biggest friction in software is the edit-compile-run loop. Remove the laptop and the context switch and you shift development from a weekend activity to a continuous, conversational flow.

This is not just productivity theater. Practitioners say the model uncovers bugs they did not know existed and can self-host parts of the stack (@steipete, @ctatedev). Early experiments like Zero: a language/tooling project that used thousands of agent tasks during prototyping: show how teams are already designing systems that assume heavy agent participation rather than treating agents as occasional helpers (@ctatedev). Read: we are moving toward closed-loop agent-driven development where the model is both co-author and test harness.

There are limits. Models still hallucinate, tests can be shallow, and a model optimizing locally may not see architecture, security, or cross-service implications. So the gains are real but conditional: they rely on robust test suites, good CI gates, and human oversight.

Performance and UX: making agents feel immediate

One practical barrier to agent workflows has been latency and UI churn. The engineering notes from @OpenAIDevs show deliberate optimization: roughly 75% less re-rendering when switching threads, removal of unnecessary re-renders on streaming paths, and substantial reductions in expensive Git operations (10x to 50x depending on the operation). The result is a more fluid, less noisy coding experience that makes long, interactive sessions tolerable.

Why does that matter beyond polish? Agent-driven cycles are conversational and stateful. If the tool behaves like a sluggish editor, you lose the speed advantage of the model. Faster reactivity enables more speculative prompts, more aggressive refactoring, and a smoother pairing session with the model. That is why performance wins are not trivial aesthetics; they are a precondition for the new workflow to scale beyond toy projects.

Security plumbing: credential brokering and the risk calculus

If you let agents act on behalf of users, you must solve credentials. Credential brokering is the practical design pattern being proposed: agents get delegated, ephemeral access to services without ever seeing raw credentials (@dangtony98). Think of it as a gateway that translates an agent's action into service calls under tightly scoped, auditable tokens. That design removes the simple exfiltration vector: you cannot leak what you never possessed.

Credential brokering is almost a prerequisite for enterprise adoption. It lets enterprises adopt agentic automation while meeting basic least-privilege and audit constraints. But it is not a panacea. Brokers centralize control, become attractive attack surfaces, and introduce policy complexity. They also force choices about who writes the policy, how fine-grained it is, and how you revoke access quickly when things go sideways. Expect a race between richer brokering protocols and new classes of misconfiguration or misuse.

There is a broader point: agents will expand the attack surface not by new cryptography but by the blunt fact that software will be asked to act more. Credential brokering is sensible mitigation; the market will build it fast if it wants enterprise dollars.

Memory, scale, and the long-lived agent

One of the persistent engineering problems for agents is history. Large models have finite context windows and retrieval systems can be noisy. The Lossless idea: compacting conversations into indexed blocks and building a tree for lookup: is a pragmatic attempt to give agents an effectively unbounded memory without losing the ability to reference precise past messages (@steipete). That model of compact + index is different from naive retrieval: it is a structural memory optimized for lookup patterns that software teams actually need.

If it works, Lossless-like architectures make agents practical collaborators on long-lived projects. They let models recall design decisions, reproduce bug histories, and maintain project-specific style. The read is that memory engineering will be as important as model quality for real-world agent usage. Expect more effort on compact representations, incremental snapshots, and provable freshness guarantees.

The labor question and ecosystem bets

Responses have split between excitement and existential dread. Some view these advances as a large experiment in the future of work: a new industrial process that shifts the value to a smaller number of higher-skill roles like agent wranglers, infra engineers, and security gatekeepers (@xeophon, @badlogicgames). Others joke about leaving tech for trades, but that quip masks a real reallocation risk.

Two ecosystem plays are emerging now. One is new tooling and languages designed to be agent-friendly (Zero being an explicit bet in that direction) that optimize for fast builds, small bundles, and predictable agent behavior (@ctatedev). The other is operational controls: brokering, audit logs, and compact memories that help organizations trust agents with more responsibility. Whoever captures these layers: the agent-friendly runtime and the safety/ops layer: will wield outsized influence over the next generation of developer tooling.

There will also be politics. As Beff noted, there are narratives around GPU supply and influence that will shape how resources and opinion leaders align on these tools. Expect debates about concentration, access, and who benefits from the productivity gains.

What to watch

• Rollout: will the ability to build and run on a paired iOS device escape a narrow beta and reach mainstream developer workflows? Watch developer adoption signals and demos of nontrivial projects built this way. (@Dimillian)
• Security standards: whether credential brokering becomes a de facto standard and whether major cloud providers publish adapters or SDKs that integrate with agent brokers. (@dangtony98)
• Memory systems: experiments that show Lossless-style compaction at scale with demonstrable retrieval fidelity and low cost. If anyone publishes benchmarks, read them closely. (@steipete)
• Agent-friendly languages and toolchains: follow Zero and similar projects for early wins or fundamental tradeoffs; note where the community opts into a new language vs. retrofitting existing ones. (@ctatedev)
• Incidents and governance: first major misuse, misconfiguration, or exploit around agent-issued actions or brokering will set the tone for enterprise adoption. Conversely, early wins in secure automation inside regulated orgs will catalyze procurement.

Bottom line: we are watching the first credible movements toward agent-native software development. The tech is still brittle and policy lags, but the ergonomics and safety primitives that make this usable are appearing in the same week. That is the signal you should be following.

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