Ramsay Research Agent — June 22, 2026

Ramsay Research Agent — June 22, 2026

Two themes ran through everything today. Models are converging, which pushes the real edge down into the harness and the prompt layer. And the open/local stack finally has weight behind it. A 3B model claiming frontier-reasoner parity dropped the same week a cluster of bring-your-own-model terminal agents crossed into "actually usable." Below: the five that mattered, the deep dives by section, and ten things you can act on today.

Top 5 Stories Today

A 3-billion-parameter model that fits in about 6.7GB of VRAM, claiming it goes toe-to-toe with a 671B reasoner. That's the headline, and I'm skeptical, and I still think it's the biggest thing today.

WeiboAI, Weibo's AI lab, open-weighted VibeThinker-3B, a dense 3B model fine-tuned from Qwen2.5-Coder-3B and trained with a post-training recipe they call Spectrum-to-Signal. The claimed numbers are loud: 94.3 on AIME26, 80.2 Pass@1 on LiveCodeBench v6, 96.1% on LeetCode. If those hold, that's parity with trillion-parameter reasoners on a model you can run on a single consumer GPU. Weights are MIT-licensed, the technical report is at arXiv 2606.16140, and it pulled 118 likes on Hugging Face inside 24 hours with VentureBeat-level coverage already poking at the benchmark claims.

Here's my problem with it, and why I'm still excited. AIME-style benchmarks are contamination magnets. A 3B model hitting 94.3 on competition math while a 671B model sits near the same number isn't "small model caught up." It's "something about this evaluation is leaking, or the training set saw the shape of these problems." I've been burned enough times by leaderboard numbers that don't survive contact with my own held-out tasks. So I don't trust 94.3. But I don't need to. Even a fraction of that capability in a 3B MIT-licensed package changes what you can put on-device, in CI, in a tight inference loop, without an API key or a per-token meter running.

What builders should do: pull it, but evaluate it on your own private tasks, not the published benchmarks. The high-value use isn't replacing your production reasoning model. It's the offline work you currently overpay for: routing, triage, classification, LLM-as-judge eval grading in CI, first-pass drafts you'll rerank anyway. A reasoner that runs free on hardware you already own is a budget line you can delete. This connects directly to the BYO-model pattern in story 5 below, and to the Unsloth quantization work in the Models section. The local stack just got a model worth bringing.

Persuasion stopped being a hypothetical risk today and became a measured number. A big one.

A four-experiment preregistered study (arXiv 2606.16475, also the lead item in Jack Clark's Import AI #462) ran 18,978 conversations across 6,923 people and found frontier AI more persuasive than expert humans. Not amateurs. Elite competitive debaters and professional political canvassers who researched the topic, practiced live, and had a £1,000 bonus on the line. The AI beat them. In a real-money test, it was roughly 3x better than a UK fundraising firm at soliciting actual Save the Children donations. The kicker: the edge held even after the human experts trained specifically against the model that had beaten them. Parity only showed up when the AI was throttled down to human typing speed and human message length.

That last detail is the one I can't stop thinking about. The model's advantage wasn't some unbeatable argument. A big chunk of it was throughput. It writes faster and longer, and at machine speed that compounds into persuasion the way compound interest compounds into money. Coach a human against it and the human still loses, because they can't match the volume.

Why it matters: if you ship anything that talks to users in an open-ended way, support agents, sales agents, onboarding flows, you are now operating a persuasion system whether you designed one or not. The capability is in the base model. You don't opt into it. You opt out of it, deliberately, with constraints.

What builders should do: treat conversational agents that influence user decisions as a regulated surface even if no regulator is watching yet. Disclose that it's AI. Rate-limit message length and turn count in flows where the user is deciding to spend money or share data. Log the conversations. And if you're building in a domain where persuasion is the product (fundraising, sales, political), assume governance is coming and build the audit trail now, not after.

The DeepMind talent drain went from rumor to pattern in one week, and the direction tells you something.

John Jumper, the DeepMind VP who co-created AlphaFold and won the 2024 Nobel Prize in Chemistry for it, announced June 19–20 that he's leaving Google DeepMind after nearly nine years to join Anthropic. TechCrunch, Bloomberg, and CNBC all confirmed it. The move lines up with Anthropic's expanding push into life sciences and computational biology, and it lands while the company is in its export-control fight with the U.S. government.

Now stack the second data point on top. Days earlier, Gemini co-lead Noam Shazeer said he'd leave for IPO-bound OpenAI, per CNBC. Google reportedly paid around $2.7B to bring Shazeer back from Character.AI less than two years ago. So in a single week, Google lost two of its most senior AI people, one to Anthropic, one to OpenAI. A Nobel laureate and a Transformer co-author, walking to the two biggest rivals at the same time.

One departure is a personnel story. Two, in opposite directions, in one week, after a $2.7B retention spend that clearly didn't retain, is a signal about gravity. Talent goes where the work feels most alive and the upside feels most real. Right now that's evidently not the Gemini org, despite Gemini being genuinely good.

What builders should do: this isn't actionable in the "change your code today" sense, but it's a real input to where you place longer bets. Jumper at Anthropic means their bio/life-sciences capability is about to get serious. If you're anywhere near drug discovery, protein work, or computational biology tooling, watch what ships from Anthropic over the next year. And if you're choosing a frontier provider to build on, talent flow is a leading indicator that the public benchmarks lag.

The single most actionable thing today: go delete "think step by step" from your prompts. It's not helping anymore. On reasoning models, it's actively hurting.

The 2026 context-engineering guidance from The AI Corner makes the case directly. Explicit chain-of-thought instructions, the ones that genuinely improved older models, are now redundant or counterproductive on reasoning-tier models that already do internal CoT. You're telling a model to do a thing it already does, and the instruction competes with the actual work for attention. The fix is to strip those directives and reinvest the freed prompt budget into context: what the model can see, retrieve, and remember at the moment it acts.

I believe this one because I've watched it happen in my own projects. Prompt templates I wrote in 2024 carry cargo-cult scaffolding I never re-tested. "Let's think step by step," "take a deep breath," role-play preambles. On the older models those moves bought real accuracy. On current reasoning models they're noise at best. The honest move is to re-audit every template you've been dragging forward and A/B it against a stripped version.

Why it matters beyond the one tip: this is the same story as model convergence. When VibeThinker-3B and a 671B model land near the same benchmark number, the model stops being your differentiator. What's left is the harness, the retrieval, the verification loop, the context you assemble. Latent Space made the same argument this week about "harness engineering" being the real coding-agent edge (more in the Agents section). The prompt layer and the scaffold layer are where the remaining advantage lives, because everyone has access to roughly the same intelligence now.

What builders should do: audit your prompt templates this week. Remove CoT directives aimed at reasoning models, measure on a held-out set, and pour the saved tokens into better retrieval and memory. The technique that worked in 2024 is now a regression you're shipping on purpose.

A bunch of separate repos quietly agreed on the same thesis this month: stop renting your intelligence.

Across June, a cluster of terminal coding agents converged on local-first, bring-your-own-model. nanocoder is community-built and model-agnostic. DeepSeek-Reasonix is DeepSeek-native and tuned for prefix caching. AionUi positions as a local 24/7 cockpit. And Mintplex's anything-llm, sitting at roughly 61.9K stars, ships under the literal banner "stop renting your intelligence." None of these is a blockbuster launch on its own. Together they're a track.

The pitch is the inverse of the hosted mainstream (Claude Code, Cursor, Codex): keep the code on your machine, point it at whatever model you want, and owe nothing to a vendor. No per-seat pricing. No code egress to someone else's servers. No model swapped out from under you on a Tuesday.

I use Claude Code every day in my personal projects, and I'm not abandoning it. The hosted tools are still better at the things that matter most: the harness quality, the tool integration, the polish. But the gap is the point. A year ago, "run your coding agent fully local" meant accepting a large capability hit. Now, with models like VibeThinker-3B and the Unsloth quantizations bringing real reasoning to a single rig, the BYO option is a genuine alternative for a specific job, not a toy. Convergence at the model layer (story 1, story 4) is exactly what makes this viable. When the model isn't the moat, you can swap in a free local one and keep most of the value.

What builders should do: if you have a codebase you legally can't send to a third-party API, client work under NDA, regulated data, internal IP you're paranoid about, stand one of these up this week and run it against a quantized local model. Even if you stay on hosted tools for daily driving, having a vetted local fallback that never egresses code is worth the afternoon. The lock-in math only gets worse the longer you wait to know your exit.

Security

The most common MCP vulnerability is the "omnibus tool," and the fix is structural, not a patch. Per the MCP Server Security Best Practices 2026 guide, a single tool like execute, query, or run that takes free-form input and dispatches on a runtime action string becomes a privilege-escalation primitive. You match handlers against a fixed enum of operations, and you make dangerous requests literally unexpressible: enums over strings, bounded integers, branded IDs in the schema. Add per-tool authorization at dispatch (never let a handler check its own permissions), and read tenant IDs from verified token claims, not the request body. This is the rare security finding where the secure version is also the cleaner design.

Force confirmation on state-mutating tools after an untrusted-content read. The same guide gives a concrete prompt-injection containment rule: if a tool ingests untrusted content (scraped pages, webhook payloads, document bodies) and a state-mutating tool fires in the same turn, require explicit confirmation before it runs. Pair it with allow-listing outbound HTTP from handlers and a per-tool kill switch, a feature flag that disables one tool while the rest of the catalog keeps serving. That cuts remediation from hours to minutes. The mental model: treat every byte a tool returns as adversary-authored.

Agents

"Harness engineering" is becoming the real coding-agent differentiator. Latent Space's AINews flags the interaction-and-verification loop wrapped around a model as the emerging edge, noting DeepSeek has stood up a dedicated harness team while Google (Gemini Managed Agents) and LangChain are formalizing the concept. The takeaway maps cleanly onto today's convergence story: model choice is flattening, so the scaffold around the model is where advantage survives. If you run a self-improving pipeline, this is the layer to invest in.

WorldLines benchmarks long-horizon, stateful embodied agents. arXiv 2606.18847 introduces a benchmark and models for agents that must carry state across many steps instead of acting one-shot, targeting the durability-of-state gap most agent evals skip. It's aimed at realistic, hours-long tasks rather than single-turn actions. If you're building agents that run for hours and quietly lose the plot halfway through, this is the eval that measures the failure you're feeling.

GateMem measures memory governance in multi-principal shared-memory agent systems. Surfaced on Hugging Face Daily Papers, GateMem benchmarks how multiple agents and users safely read and write a shared memory store without leaking to or clobbering each other. That's the access-control and isolation problem nobody measured while everyone raced to ship multi-agent setups. If you run an agent fleet against one memory backend, this is the eval to watch before a cross-tenant leak finds you first.

Thorsten Ball: "a lot of code will be written while nobody is watching." In his latest Register Spill, the Sourcegraph/Amp engineer argues the near future is agents running longer and everywhere, kicked off from anywhere, with code increasingly written unattended. He points to Google's new gws ("CLI for all of Google Workspace, built for humans and AI agents") and Simon Willison's growing Agentic Engineering Patterns project as concrete signals. Single-source commentary, but a sharp read on where the workflow is going.

Research

G2Rec unifies graph co-engagement modeling with semantic tokenization for generative recommendation. A Meta-affiliated team introduces G2Rec, which captures user interest patterns without ground-truth interest labels and reports online deployment across production surfaces plus public-dataset gains over prior generative-rec and graph-tokenization baselines. The interesting bit for practitioners: it goes after the two things that actually block recsys deployments, the scalability ceiling of graph methods and the heuristic, unsupervised nature of semantic tokenization. Worth a read if you're anywhere near large-scale recommendation.

"The token is a group element": Lie-algebra attention over matrix Lie groups. Przemyslaw Musialski reframes a token as an element of a matrix Lie group instead of a feature vector, scoring attention via a parameter-free negative squared algebra norm of the relative pose. On sequence-completion over SE(2), SO(3), and Aff(2) it matches learned MLP kernels with 50–80x fewer score parameters, while vector-token baselines violate invariance by 5–12 orders of magnitude. It extends equivariant attention to non-compact, non-abelian affine groups earlier methods couldn't handle. Niche, but directly relevant to robotics, pose, and geometric ML.

First cross-attention attribution study for style-captioned TTS. Mathur, Sayed, Madha et al. adapt the DAAM framework to speech diffusion, analyzing 3,600 style-caption/transcript pairs across 25 layers and 24 ODE steps in CapSpeech-TTS. They find style tokens have lower temporal variance than content tokens (confirming global conditioning), correlate with F0 and energy, and peak in early generation steps and deeper layers, with attention entropy bottoming at layer 17 where style importance peaks. Narrow on immediate deployment, useful for anyone doing controllability work in instruction-driven TTS.

ByteDance's PerceptionDLM brings parallel region perception to multimodal diffusion LMs. Per Hugging Face Daily Papers, PerceptionDLM processes multiple visual regions in parallel using diffusion-based language models instead of autoregressive decoding. It's another data point that diffusion-LM architectures are getting real traction in vision-language, where parallelism can cut latency on complex scenes. One to watch if you care where multimodal efficiency is heading, especially for dense, many-object images.

Infrastructure & Architecture

MCP published the release candidate for its 2026-07-28 spec, the biggest revision since launch. The MCP blog calls it the largest revision yet, with the final spec due July 28. Governance has moved to the Linux Foundation, Streamable HTTP has made remote MCP mainstream, and there are now 14,000+ servers available. The RC window is your chance to test client and server changes before the spec freezes. If you maintain an MCP integration, test against the RC now, not after July 28 when the surprises become production incidents.

Datasette Apps: sandboxed hosting for AI-generated HTML apps next to your data. Simon Willison shipped a Datasette plugin that hosts custom HTML applications inside Datasette, with CSP-based sandboxing aimed specifically at safely running AI-generated apps. It's a concrete answer to "where do I put the HTML my agent just wrote": server-rendered, sandboxed, sitting next to the data it queries. A clean pattern for shipping agent-authored micro-apps without standing up new infrastructure for each one.

Tools & Developer Experience

Cursor 3.7 adds /in-cloud for cloud subagents on dedicated VMs with reusable snapshots. Cursor's June 17 release launches subagents on dedicated VMs, environment setup completing in under 10 minutes, with snapshots to speed future runs. Combined with the new Automations layer, it gives solo builders a way to offload long-running or parallel agent work off the laptop without hand-managing infrastructure. The snapshot reuse is the real win, since the first-run setup cost amortizes to near-zero on repeat tasks.

Claude Code v2.1.185 softens the stream-stall hint. The June 21 release reworks the mid-stream stall notice to read "Waiting for API response · will retry in …" instead of the alarming "No response from API · Retrying in …," and only after 20 seconds of silence rather than 10. Tiny change, real effect: it stops triggering false-alarm anxiety during long reasoning runs where brief silences are normal. The kind of UX detail that only matters if you live in the tool, which is exactly why it matters.

Simon Willison ships sqlite-utils 4.0rc1 with schema migrations and nested transactions. The first major version bump of his widely-used Python/CLI SQLite toolkit lands built-in schema migrations and savepoint-based nested transactions on June 21. If you lean on sqlite-utils as glue in data or LLM pipelines, migrations kill the hand-rolled ALTER scripts you've been maintaining. It ships the same week as his Datasette Apps work, continuing the push to make SQLite a first-class substrate for agent-built tools.

Claude Code's June updates doubled rate limits and added nested skills. Anthropic's June changelog doubled Claude Code rate limits, raised Opus API limits, and shipped nested skills, streamlined agent teams, and tighter permissions and auto-mode review. The headroom and nested-skills support matter most if you run Claude Code heavily in automated agentic loops, where you were previously hitting the ceiling. Distinct from the Artifacts work covered earlier.

OpenAI adds credit usage analytics and spend controls to ChatGPT Enterprise. Announced around June 18–20, the Global Admin Console now breaks down ChatGPT and Codex credit consumption by user, product, and model, with workspace defaults, per-group limits, and individual overrides. It targets CFOs staring at escalating per-seat AI bills from power users. Available to enterprise customers immediately. The interesting subtext: AI spend is now unpredictable enough that vendors ship the cost-control tooling before customers ask.

Models

Unsloth squeezes GLM-5.2's 1.51TB down to 217GB at 1-bit, plus ~2x-faster Gemma 4 MTP. Unsloth's changelog shows GGUF quantizations compressing GLM-5.2's 1.51TB BF16 weights to 217GB at 1-bit, bringing the strongest text-only open model into reach of a single high-memory local rig. The same update adds Gemma 4 MTP with auto speculative decoding for roughly 2x faster inference. For builders who want frontier-open quality without an API dependency, this is the practical "can I actually run it" answer, and it's the same local-first thread running through today's Top 5.

Vibe Coding

Windsurf rebrands to Devin Desktop. Per AIToolTier, Windsurf has folded into the Devin agent brand as Devin Desktop, shifting from editor-with-agents toward an agent-first desktop product. More consolidation in the AI-IDE space. If you standardized workflows on Windsurf, expect branding, pricing, and feature realignment. Worth watching whether the editor experience survives the agent-first repositioning or gets subordinated to it.

An Ask HN thread maps the converging default AI dev stack, and a more skeptical mood. A widely-discussed "What is your AI dev tech stack/workflow?" thread surfaces the now-standard profile: Python/Go/TypeScript plus coding agents, vector stores, and LLM-API glue. Alongside it, a notably more skeptical tone about whether agent-generated code creates more cleanup than value. Good ground-truth on how working engineers actually wire AI into their stacks right now, versus how vendors describe it. The skepticism is the signal worth tracking.

SaaS Disruption

The agent layer is being bought, not built. SaaStr catalogs a single week of M&A where the buyer chose acquisition over organic build to own the agent surface: Salesforce bought Fin/Intercom ($3.6B, support), SpaceX bought Cursor ($60B, devtools), and Asana acquired no-code agent-orchestration platform StackAI. Three unrelated categories, same move. As AI-native challengers hit distribution scale, the cheapest path to "system of record for agents" is a checkbook, not a roadmap. Expect more category leaders to flip from defenders into acquirers.

The AI SDR category splits into seats versus outcomes. Vector Agents' 2026 comparison shows a clean bifurcation. AI-assisted incumbents still price per human (Apollo $49–99/seat, Outreach $100–160/seat) while fully autonomous agents price per outcome or per machine: 11x.ai at $5,000–$10,000/month, AiSDR around $900/month, Artisan/Ava from $280/month. The split is the clearest signal yet that the seat breaks down once the agent, not the rep, runs the sequence, and buyers are being asked to pay for booked meetings instead of logins.

New martech launches collapse the whole stack into one autonomous layer. Per MarTech, Ploy.ai targets full-funnel growth with autonomous landing pages, SEO, ABM, and ad attribution, while Elentaria handles B2B channel selection and sequenced GTM execution. Incumbents bolt AI on top (Jasper became a Content Automation platform, Intuit Mailchimp added natural-language analytics), but the new entrants are architected to replace the bundle, not augment it. The pattern: point tools are getting absorbed into operators that run the whole funnel.

Salesforce raises list prices ~6% while bundling "unlimited" Agentforce. Salesforce Ben reports the price hike paired with unlimited Agentforce licenses on certain tiers, effectively folding agent usage into the seat instead of metering it. It's a strategic bet that "agents are free with the platform" beats per-agent metering for lock-in, even as Agentforce's à la carte consumption pricing persists elsewhere. The live tension: are AI agents a line item or a platform entitlement? Salesforce just voted entitlement.

CRM forks into rip-and-replace versus overlay. SaaStr frames the AI-CRM market as splitting on architecture, not features. Lightfield and peers (Attio, Monaco, Reevo) pursue full rip-and-replace of Salesforce/HubSpot with self-updating, memory-native systems, while Aurasell layers an AI-native go-to-market OS on top of existing CRMs with no replacement required. The reframed buyer question is "follow the agents": pick the platform where your agents do the most work. Agent capability, not UI or integrations, becomes the selection criterion.

Policy & Governance

The clearest rationale yet for the Fable 5 / Mythos export ban: the capability is the concern, not a patchable exploit. As of June 21, Tech Policy Press surfaces The Economist's reporting that the models' autonomous offensive-cybersecurity capability is itself the national-security issue, not a user-side exploit you can fix. That reframes the dispute from "why banned" to a structural capability problem, which is a much harder thing to negotiate away. Anthropic's International MD Chris Ciauri said the models should return "in the coming days," and prediction markets give 57% odds of restoration before July 1, 75% by July 17.

New alignment org "Sequent" launches on the thesis that alignment is not on track. Jack Clark's Import AI #461 covers Sequent, founded on the view that frontier labs' reactive approaches won't keep pace with superintelligence development. Instead of reacting, it plans a portfolio of "differentiated alignment bets." Worth noting for anyone tracking where independent safety capital and talent are flowing, especially in the same week two Nobel-and-Transformer-tier researchers changed employers.

Skills of the Day

1. Default RAG to hybrid retrieval plus reranking. Run dense embeddings for meaning and sparse/BM25 for exact terms, part numbers, and rare entities, then a cross-encoder rerank pass. That's a 25–40% precision lift over naive RAG at modest latency, per Starmorph. If you only have dense search today, adding the sparse channel is the highest-leverage upgrade you'll make this month.

2. Truncate embeddings with Matryoshka instead of re-embedding. Matryoshka representations (OpenAI, Cohere, Jina) let you cut a vector to fewer dims and keep most retrieval quality. Store full vectors, serve a 256–512 dim prefix for first-pass search, then rerank at full dimensions. Ailog's benchmark puts most apps best at 768–1024 dims. It's a near-free latency and storage knob most teams leave on max.

3. Drive overnight agent runs off a JSON feature file. Use a structured JSON list (each feature "passes": false, JSON over Markdown so the model won't casually rewrite it), a claude-progress.txt log, and an init.sh that boots the dev server and runs end-to-end checks before any coding. Anthropic's harness pattern has each session read git logs and progress first, so a fresh run refuses to build on a broken state. Built for daily-cron agents.

4. Pin Claude Code version and model in managed settings. requiredMinimumVersion/requiredMaximumVersion (v2.1.163) make Claude Code refuse to start outside an approved band, and enforceAvailableModels (v2.1.175) constrains the default to an allowlist, per the changelog. Set these in managed config so a fleet of machines can't silently drift on version or model. Stops the "works on my agent" class of bug cold.

5. Run a small local reasoner as your CI eval judge. Stand up a quantized 3B-class reasoner on hardware you own and use it for LLM-as-judge eval grading in CI. Keep your paid API calls for production traffic, not the test loop that runs on every push. The math: eval grading is high-volume and low-stakes, exactly the workload a free local model should eat.

6. Make dangerous MCP operations unexpressible in the schema. Don't ship an execute tool that dispatches on a runtime string. Use a closed enum of operations, bounded integers, and branded IDs so a privilege-escalation request literally can't be constructed. Read tenant ID from verified token claims, never the request body. The type system is your cheapest security control.

7. Give every MCP tool a per-tool kill switch. Feature-flag each tool so you can disable exactly one compromised tool in seconds while the rest of the catalog keeps serving. That turns a multi-hour incident into a one-line config change. Build it before you need it, because you'll need it during the incident, not after.

8. Load tool schemas on demand instead of front-loading dozens. Stuffing 50 tool definitions into every context window costs tokens and raises wrong-tool selection errors. Search and fetch schemas as needed. Smaller context, sharper tool choice, lower bill. The same logic as lazy-loading anything else: don't pay for what this turn won't use.

9. Use savepoints for multi-step schema migrations. Nested transactions let a failed step in a migration roll back cleanly instead of leaving a half-migrated database you have to repair by hand at 2 a.m. sqlite-utils 4.0rc1 ships these natively, so stop hand-rolling ALTER scripts and wrap the migration in a savepoint.

10. Set per-group AI spend limits before a power user blows the budget. In your enterprise admin console, break down credit consumption by user, product, and model, then set workspace defaults and per-group caps. AI spend is now unpredictable enough that the cost-control tooling shipped before most teams turned it on. Turn it on now, not after the invoice.