Agent identity, agentic commerce, and a UK government stress-test of five AI futures

Three things that looked like edge-case problems last month — agent authentication, agent payment authorization, and agent research reliability — turned into the center of the industry on June 15. Here is what moved.

NewCore emerged from stealth on Monday with $66 million to build what its CEO Zohar Alon calls a "ground-up" identity platform for enterprises that are treating AI agents as workers rather than software. 1 The round was led by Cyberstarts, with Index Ventures and Evolution Equity Partners joining; post-money valuation is $300 million. The founders — Alon (ex-Check Point via Dome9 acquisition), CTO Amihai Neiderman (Unit 8200 research lead), and CCO Erez Yarkoni (former CIO of T-Mobile USA and Telstra) — argue that Okta and Entra were designed for humans and are being stretched past their breaking point.

NewCore's architecture uses a "split-key" model: critical identity credentials are divided between the customer and the platform, eliminating a single point of compromise. It ships an "Agentic Skill" integration for Claude Code, OpenAI Codex, and Cursor so those tools can authenticate as managed identities rather than manually distributed credentials. Employees can approve or revoke agent access via a mobile app.

The same day, Akamai announced a unified six-pillar agentic security framework built around a Know Your Agent (KYA) protocol, developed with Visa, Skyfire, and Experian. 2 KYA standardizes how agents declare identity, origin, and intent, linking them to both the platform they operate on and the authorized human they represent. The framework's commercial hook is explicit: verified agent identity is the precondition for agents participating in payment flows, and Akamai is positioning its edge network as the runtime enforcement layer.

A third announcement from Trust3 AI — rebranded from Privacera — launched AgentDOS, described as the first enterprise control plane for agent observability. 3 Its pitch is token-consumption monitoring: a healthcare customer deploying agents on Azure and Databricks found, within days of deploying AgentDOS, that two agents were accessing regulated patient datasets without valid purpose context, and a third was on track to exhaust its monthly token budget in 11 days.

Taken together, these three announcements define the same architectural gap: agents exist inside enterprise networks today but have no reliable way to prove who they are, what they're allowed to touch, or how much they should cost. The identity problem has moved from design-time debate into funded product.

Visa and OpenAI announced a partnership at the Visa Payments Forum in San Francisco on June 15 to integrate Visa's network, tokenization, and fraud infrastructure directly into OpenAI agentic experiences. 4 Transactions will run inside user-defined policies — spending limits, merchant category filters, required approvals — using tokenized Visa credentials with real-time authorization and monitoring.

This is the same infrastructure layer Akamai is building around from the security side. Visa already has Skyfire and Experian in its KYA ecosystem via the Akamai deal; the OpenAI partnership adds the demand side: Codex and other OpenAI developer tools will be able to initiate Visa transactions on behalf of authorized users. Jack Forestell, Visa's Chief Product and Strategy Officer, said the company is positioning itself explicitly for a world where AI agents are the primary interface for commerce — not a complement to existing interfaces.

For practitioners building consumer-facing agents: this is the formal announcement that a major payment network has picked a side. Agentic commerce is no longer an R&D concept; it now has network rails, tokenization infrastructure, and enterprise-grade fraud monitoring.

Sakana AI launched its first commercial product on June 15 — Sakana Marlin, a long-horizon research agent that runs autonomously for up to roughly eight hours, then delivers a detailed strategy report of up to 100 pages along with executive summary slides. 5 The company describes it as a "virtual CSO" — able to handle the kind of strategic market or competitive research that would otherwise take a small team several weeks.

The technical foundation draws on two prior Sakana research projects: AB-MCTS (NeurIPS 2025 Spotlight), a multi-model coordination method, and the AI Scientist pipeline (published in Nature in 2026). The agent cycles through hypothesis formation, information gathering, and verification without human input after the initial topic and scope dialogue.

Pricing is pay-per-use at the entry tier, with Pro, Team, and Enterprise plans. Regional availability restrictions apply (the service launched primarily in Japan). Around 300 beta testers from financial institutions, consulting firms, and think tanks contributed to the refinement cycle starting April 2026.

For the agent builder audience: Marlin is the clearest evidence yet that long-horizon agentic pipelines — 8+ hours, iterative hypothesis-and-verify loops, structured outputs — are production-ready with current foundation models. The constraint is no longer capability; it is cost modeling, session stability, and output quality assurance for high-stakes professional deliverables.

The UK Government Office for Science published AI Scenarios 2030 on June 15, an updated policy planning framework co-developed with the AI Security Institute (AISI) and the Department for Science, Innovation and Technology. 6 The report lays out five scenarios built around six critical uncertainties: capability trajectory, distribution and model access, security controllability, adoption breadth, labour displacement, and global cooperation.

The five scenarios break across three capability trajectories:

The document explicitly notes that as of 2026, AI systems "already operate with high autonomy and surpass experts in certain domains," and that even in the slowdown scenario, capability gains continue through better integration and productization. The frontier market is expected to remain concentrated around a small number of US and Chinese companies through 2030, with most countries competing for access rather than frontier position.

For anyone stress-testing product or investment strategy: the report is unusual in that it was designed specifically for government policy workshopping, not public communication, which makes it unusually concrete about failure modes. The Take-Off scenario — where safety is deprioritized amid race dynamics — is the one most directly relevant to the ongoing Fable 5/Mythos 5 controversy.

Parallel-Synthesis (arXiv:2606.14672, submitted June 12) proposes replacing the standard "concatenate all branch outputs as text" synthesis step in multi-agent workflows with a direct KV-cache merge. 7 The paper targets the common pattern where independent agents explore subtasks, retrieve evidence, or generate candidate solutions in parallel before a synthesizer combines them.

The problem with text concatenation: it discards the parallel structure, adds redundant prefill computation, and forces the synthesizer to re-process information it already processed in the worker context. Parallel-Synthesis uses a cache mapper to calibrate independently generated branch caches and a fine-tuned synthesizer adapter that generates directly from the merged non-sequential cache.

Results across nine datasets (math, science QA, code generation, GAIA, multi-agent database diagnosis): matches or outperforms text-based synthesis on seven of nine, and reduces time-to-first-token by 2.5×–11×. The framework is designed as plug-and-play, requiring fine-tuning the adapter but not retraining the base synthesizer.

Practical implication: if your agent pipeline has parallel branches feeding a central synthesizer — common in research agents, multi-tool orchestration, and debate-style reasoning — this paper offers a concrete path to dramatically lower latency without sacrificing synthesis quality. The 2.5×–11× TTFT range is wide and dataset-dependent, but even the floor is meaningful at production scale.

GLM-5.2, 1M-token context, no benchmarks. Z.ai shipped GLM-5.2 on June 13 (reported June 15) with a 1,000,000-token context window (glm-5.2[1m]) and two thinking-effort levels (High and Max). 8 No SWE-bench or Terminal-Bench results at launch — the release focused on availability and the open-weights roadmap (MIT license, weights pending the following week). It is compatible with Claude Code, Cline, and OpenClaw via Anthropic-compatible endpoint swap, making it a drop-in fallback for teams affected by the Fable 5 access disruption.

The Perplexity self-study problem. A paper circulating since June 8 (arXiv:2606.07489) claims Perplexity Computer cuts task time by 87% and cost by 94% versus its answer engine. Three of four authors work at Perplexity; the only products compared are Perplexity's own. 9 The matched-pair methodology is defensible; the conflict of interest is not. A November 2025 CMU/Stanford study found AI agents completing tasks faster but with significant quality gaps, and reported that verification overhead can slow human work by 17.7%. Vendor efficiency figures are migrating into pitch decks and board memos across the industry — this is a good week to ask about independent replication before citing the 94% number.

Stratechery on Anthropic's safety position. Ben Thompson's June 15 essay argues that Anthropic's willingness to challenge the US government's Fable 5 export restriction is not inconsistency — it is the logical result of the company's belief that its safety practices give it license to push aggressively on commercial deployment. 10 Worth reading alongside the ongoing Washington standoff: Anthropic staff were in DC on June 14 seeking to characterize the government action as a misunderstanding; White House officials were characterizing the company's posture as insouciance about national security. Thompson's frame — that the conflict was structurally inevitable — is the more durable analytical lens.