The Wake: May 7, 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 6, 2026. Signals: 23 bookmarks and 2 likes.

Brief

The race in AI is shifting out of the lab and into systems: networking protocols, agent orchestration, memory, billing, and developer ergonomics. Over the last 24 hours you saw releases and moves that are less about headline model capabilities and more about making agents reliable, scalable, and commercially viable. OpenAI pushed an open networking spec to reduce wasted GPU time. Anthropic and partners are betting on better agent infrastructure to make long-horizon agents useful. Tooling vendors are racing to handle state, resumability, billing, and code search. Expect the next phase of competition to be decided by systems engineering, partnerships, and product design rather than raw model size.

Infrastructure is now the product

OpenAI announced Multipath Reliable Connection, an open protocol for running large training clusters with less wasted GPU time, and named partners including AMD, Intel, Broadcom, Microsoft, and NVIDIA. This is a clear signal: optimization at the network and systems layer is an independent lever to reduce cost and improve throughput. It also signals a shift toward open, interoperable infrastructure rather than proprietary stacks that lock partners in.

Why it matters: raw model quality matters less if you cannot effectively train or orchestrate models at scale. Anyone trying to compete on model throughput or cost-per-token now has a standards-level target to hit. Expect vendors and cloud providers to either adopt MRC or announce compatible alternatives quickly.

Agents moving from demo to delivery

Anthropic just launched long-horizon agents, and their partners are already positioning to provide the orchestration and memory layers those agents require. That was the explicit read in public congratulatory notes and partner commentary. Long-horizon agents are different from chat models: they need persistent memory, reliable state restoration, and coordinated multi-agent workflows to handle complex tasks over time.

The practical challenges are visible in the ecosystem: research from Entire shows that almost half of agent tool calls involve reading and searching code and files, and that the bottleneck is not raw search speed but how agents reason about which files matter. That squares with Anthropic’s and partners’ messaging that agent infrastructure is more than compute: memory management, indexing, and multi-agent choreography are critical.

If you are building an agent product, three engineering priorities matter now: (1) durable session state and memory primitives, (2) robust retrieval/indexing that surfaces relevant assets, and (3) orchestration and observability so agents don’t silently go off the rails. Vendors are racing to supply those primitives.

Tooling and developer ergonomics are catching up

You saw multiple smaller but important signals about the developer experience. Claude is raising rate limits and removing peak-hour throttles for Code-focused plans, which is a classic volume play: make the product more usable, reduce friction, and capture developer mindshare. Tools like cmux that restore Claude sessions across restarts and new resumable-stream replacements that avoid Redis/Upstash remove operational friction for teams building agents into production.

Other utility launches and updates: resumable streams, Polar for usage billing, better local diff tools, and Twilio’s next-generation platform: are all making it easier to run, measure, and monetize AI features. The throughput story is clear: as systems stabilize, business models and billing primitives must be ready to monetize fairly and reliably. Polar’s pitch is effectively that charging for inference can be operationalized in three steps; whether that simplifies or fragments billing approaches will depend on adoption and standards.

The macro argument: jobs, adoption, and the social conversation

The tenor from industry and investors is mixed. Dario and Daniela Amodei’s remarks re-emphasize an exponential narrative while acknowledging real social tradeoffs like unemployment. At the same time, a16z pushes back on apocalypse narratives with labor data showing software roles and related wages still growing in many AI-exposed sectors.

Read: the debate is shifting from “AI takes all jobs” rhetoric to a more nuanced argument about skill shifts and labor market churn. That helps explain vendor behavior: they are betting that productized agents and developer-friendly tooling will create new software primitives and job types even as some existing roles are automated. The political and regulatory fallout is not guaranteed, but public sentiment and storytelling will shape hiring, adoption, and policy in the next 12 to 24 months.

Tactical implications for product and procurement

• If you run an LLM or agent program, prioritize systems architecture as much as model selection. The marginal gains from better networking, retrieval, and orchestration will compound.
• Negotiate infrastructure partnerships now. Open protocols like MRC create the opportunity to standardize across vendors but also the risk of vendor lock-in if you accept proprietary alternatives.
• Reduce single-point failure in state and session handling. Tools that persist sessions and provide resumability are low-hanging fruit for improving reliability and user trust.
• Re-evaluate rate limits and quota strategy. Anthropic’s decision to loosen Code quotas and raise API limits is a play for developer mindshare; competitors should model the revenue and cost tradeoffs of similar moves.
• Measure downstream success beyond latency and perplexity. Track task completion, agent hallucination rates, and multi-step reliability. That is what customers will pay for when agents move from experiments to business-critical processes.

What to watch

• OpenAI’s MRC adoption. Which cloud and hardware partners adopt the protocol first, and whether it materially reduces GPU waste in published benchmarks.
• Anthropic’s agent rollouts and partner integrations. Look for products that expose memory primitives and multi-agent orchestration as first-class APIs.
• Developer billing and rate strategies. Watch Claude’s usage after the limit changes and Polar’s billing adoption to see whether simpler billing increases churn or retention.
• Research into retrieval and agentic search from groups like Entire. Improvements here will directly impact agent latency and correctness.
• Twilio’s platform announcement and Digg’s AI news alpha. Both are signals that adjacent markets are enabling faster integration and discovery for AI features.

Short version: the new battlefield is systems. Winning will require stitching networks, memory, orchestration, and billing into developer-friendly products that actually complete tasks over time. Expect the next wave of winners to be the teams that treat AI as a distributed systems problem first and a modeling problem second.

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