Open Weights, Code Review, and Agent Identity

A week after a US order pulled Anthropic's Fable 5 and Mythos 5 offline for foreign nationals, buyers moved to reduce their reliance on any one provider. The curbs drove two responses directly: Zhipu open-sourced GLM-5.2 as its Hong Kong shares jumped, and European firms began spreading work across providers, Reuters reports. Separately, and timed to the same week, OpenRouter shipped Fusion, which it says matches Fable 5-level quality at about half the cost by combining cheaper models.

The detail: OpenRouter's Fusion, launched on 12 June, sends a prompt to a panel of models in parallel and has a judge model merge their best answer. On Perplexity's 100-task DRACO research benchmark, OpenRouter says a budget panel of Gemini 3 Flash, Kimi K2.6 and DeepSeek V4 Pro, synthesised by Opus 4.8, scored 64.7%, edging solo GPT-5.5 (60.0%) and Opus 4.8 (58.8%) for about half the cost. It sits just behind solo Fable 5 at 65.3%, though that figure covers only the 93 of 100 tasks Fable's content filters allowed. OpenRouter is candid that Fusion is no drop-in: it runs 2-3x slower, and Fable still leads on the long-horizon work the benchmark skipped. Zhipu, now Z.ai, released GLM-5.2 under an MIT licence, a 753-billion-parameter open-weights model (Mixture-of-Experts, about 40B active per token) with a 1-million-token context aimed at long-horizon coding. Its shares spiked as much as 48% intraday before closing up about 33%, as the firm cast the release, by CNBC's account, as a reply to Washington: cutting-edge intelligence should be open and not "withdrawn at any time."

Related: It extends a run of open-weight releases Pulse24 has tracked, from Hugging Face's open reproduction of DeepSeek-R1 to Chinese models now leading global token use on price; the same week, DeepSeek raised $7.4bn at a $50bn valuation. One finding cuts the other way: one study argues open weights can widen the global capability gap rather than close it, since the compute to fine-tune and serve them stays concentrated.

A widely shared engineer's post and an academic teardown landed the same week on the same point from different angles: the hard part of AI coding is no longer writing the code. Developer Vinicius Brasil rejects AI code even when it works, arguing review is the new bottleneck, while a research paper estimated most of Claude Code is operational scaffolding, not model logic, putting the real engineering in the harness around the model.

The detail: Brasil's point is that AI makes implementation cheap while making review and judgment more expensive. He rejects changes he cannot explain in his own words, where the diff is bigger than the problem, or where he is trusting the output over his own understanding. The paper, "Dive into Claude Code" from researchers at Mohamed bin Zayed University of AI, analysed the tool's TypeScript source at version 2.1.88; it dates from April but resurfaced this week through Semafor. It reports that only about 1.6% of the code is AI decision logic and about 98.4% is operational harness, covering permissions, context compaction, hooks, subagent isolation, and session persistence. That split is an estimate drawn from reverse-engineered source, not an audited figure. Semafor's Reed Albergotti read it as the model taking a back seat to the orchestration layer around it, the design the authors call "minimal scaffolding, maximal operational harness."

Related: It follows GitButler's agent-built Git rewrite, which passed 99% of the test suite but which its author called not production-ready, and Stripe's "Minions", which merge more than 1,000 pull requests a week with no human-written code. The volume is real enough that Microsoft is adding AWS capacity for GitHub to absorb an AI-driven surge in coding. Pointing the same way, one analysis warns unreviewed agent output is creating "verification debt", the review burden Brasil describes scaled up.

Two releases this week let AI agents authenticate and hold identity without borrowing a human's. Cloudflare added Temporary Accounts so an agent can deploy code before any human signs in, and NewCore launched from stealth to give agents managed corporate identities.

The detail: With Cloudflare's feature, an agent runs `wrangler deploy --temporary`; Wrangler now nudges the agent toward the flag, and Cloudflare provisions a throwaway account, API token and live Worker for 60 minutes, returning a claim URL that pulls in not just the Worker but any databases and bindings it created. Unclaimed accounts are deleted, and Cloudflare says production work should still use a permanent account. It is one of several signup-removing moves, alongside a Stripe partnership and the WorkOS auth.md standard. NewCore, founded by former Dome9 chief Zohar Alon, came out of stealth with $66m at a $300m valuation, per TechCrunch, in a seed led by Evolution Equity after an Index Ventures and Cyberstarts pre-seed. It ships an "Agentic Skill" that lets coding agents such as Claude Code, Codex, and Cursor sign in as managed identities rather than on borrowed credentials, with lifecycle and revocation control, and is live with fewer than 10 customers.

Related: It runs alongside a wider build-out of agent plumbing: Google's Interactions API reached general availability with managed agents and background execution, and agents keep spreading across enterprise stacks, from data tools to Salesforce's $3.6bn purchase of the Fin support agent. The same autonomy widens the attack surface: NewCore's launch lands the month Microsoft pulled at least 70 GitHub projects after password-stealing malware was planted in AI developer tools, and a Stanford report found adoption is outpacing governance as incidents rise, the gap NewCore's revocation controls are pitched to close.

---