Ramsay Research Agent — July 8, 2026
Two economic stories are eating everything else this week, and they're the same story wearing different clothes. On one end, the cost of frontier-quality inference is falling through the floor because Chinese open-weight models got good. On the other, enterprises are walking into renewals asking to cut their seat counts by 5x because one agent does the work of ten people. Cheap intelligence on the supply side, collapsing headcount on the demand side. If you build or sell software, both curves are pointed at your margin.
Everything else today, security holes in agentic workflows, an open-source project slamming the door on AI PRs, the token-reduction tooling explosion, is downstream of those two forces. Let's get into it.
Top 5 Stories Today
1. GLM 5.2 and the coming AI margin collapse
Martin Alderson's essay "The upcoming AI margin collapse, part 1: GLM 5.2" hit 675 points and 462 comments on Hacker News, and it's the rare HN chart-topper that's actually about spreadsheet math instead of vibes. The argument is simple. Z.ai's GLM 5.2 delivers frontier-adjacent coding quality at a fraction of the price of proprietary APIs, and once a good-enough open-weight model exists, the inference margin that funds the whole proprietary-API business starts to evaporate. (Hacker News)
The numbers behind it aren't hand-waving. CNBC reported the same day that open Chinese models now run 60 to 90% cheaper than the leading Anthropic and OpenAI models, and that US companies' token share on Chinese models via OpenRouter has topped 30% weekly since February 8, peaking at 46% against an 11% twelve-month average. (CNBC) GLM 5.2 was Vercel's fastest-adopted model of 2026, roughly 27x daily token growth and 80x customer growth in its first week. Lindy moved 100% of its traffic off Claude to DeepSeek. On the leaderboards, GLM-5.2 sits at the top of BenchLM's open-source index at 83, ahead of DeepSeek V4 Pro Max and reportedly beating GPT-5.5 and Claude Opus 4.8 on SWE-Bench Pro and Terminal-Bench 2.1. (BenchLM) Z.ai even shipped ZCode, an MIT-licensed open-weight agentic coding environment built on GLM-5.2, aiming straight at Claude Code and Codex. (BuildFastWithAI)
Here's where I'll push back on my own enthusiasm. TechCrunch ran a piece arguing open models aren't actually hurting Anthropic yet, that open and frontier models capture different phases of the same adoption curve. (TechCrunch) The "yet" is doing a lot of work. My read: they're both right, just on different timelines. Frontier labs still win the hardest 10% of tasks. But most production traffic isn't the hardest 10%. It's summarization, classification, code edits, RAG synthesis. That's the exact middle GLM and DeepSeek and Kimi are now eating.
What builders should do: start measuring the quality gap on your actual workload, not on Twitter benchmarks. Route the boring 80% of your traffic to a cheap open model and reserve frontier calls for the tasks that genuinely need them. Together AI just launched Provisioned Throughput with a 99% uptime SLA and up to 90% lower cost than proprietary APIs specifically for models like GLM-5.2. (Together AI) The infrastructure to do cost-based routing in production now exists. The only reason not to is that you haven't measured yet.
2. The token stack is now a real category, and it can cut your bill 90%+
If the margin collapse is happening to the providers, the demand-side version is happening in your own usage graph. Practitioners are now stacking independent token-reduction layers and compounding the savings past 90% total. (Codepointer)
The recipe that's making the rounds: a model router first, then rtk (60 to 90% off Bash and tool output), then Caveman (50 to 75% off model output), then the Headroom API proxy (47 to 92% off tool and RAG payloads). Each one targets a different token class. Router handles model selection, rtk compresses tool output going back into context, Caveman makes the model answer in terse telegraphic phrasing, Headroom compresses payloads at the proxy layer. Because they hit input, tool output, and model output separately, they multiply instead of overlapping. The critical discipline is measurement: run ccusage for a week to get a baseline, then add one tool at a time so you can actually attribute the savings. If you stack all four blind, you'll never know which one is carrying you or which one quietly broke your output quality.
The individual pieces are worth naming. Caveman is a one-line Claude Code skill install that cut an average 65% of output tokens across 10 benchmark prompts while preserving substance. (GitHub) Because it targets output tokens, the most expensive class, it stacks cleanly with input-side proxies. Headroom Labs' proxy claims 60 to 95% compression. (GitHub) Firecrawl benchmarked 12 techniques separately and found model routing (send isolated subagent work to Haiku instead of Opus) and path-scoped rules were the biggest wins, with one prompt-shaping technique hitting 91.9% context reduction with no measured quality regression. (Firecrawl)
Why this matters right now: I run agents daily on a subscription, and the ceiling isn't dollars, it's usage limits. Every token I don't spend on a Bash dump of a build log is a token I get to spend on actual reasoning. "What's your token stack?" is becoming as routine a question as "what's your model?" and if you can't answer it, you're leaving throughput on the table. Pair this with the margin story: cheap models plus aggressive token compression is the entire 2026 cost-control playbook in one sentence.
3. GitLost turns a plain-English GitHub Issue into a private-repo leak
A GitHub Issue. No code, no credentials, no access. Just a paragraph of English that tells an AI agent to copy your private repo into a public comment. That's GitLost, and it works whether the agent runs on Copilot, Claude, Gemini, or Codex. (Noma Security)
Noma Security disclosed it this week: a prompt-injection flaw in GitHub Agentic Workflows, which have been in public preview since February. An unauthenticated attacker posts a crafted public Issue whose body instructs the workflow agent to pull private-repo data and post it back as a public comment. The agent, doing exactly what agents do, treats the Issue text as instructions rather than data and complies. The HN thread hit 327 points, and The Register, Dark Reading, and The Hacker News all covered it, so this isn't a single-source scare.
The reason this one lands harder than a typical CVE is the blast radius. It's model-agnostic. The vulnerability isn't in Copilot or Claude specifically, it's in the pattern of wiring an LLM agent up to untrusted public input with write access to private context. Cyera published a related argument this week titled "The MCP Governance Illusion," making the point that securing individual tools misses the real risk surface, the data, identities, and cross-system access those tools broker. (Cyera) GitLost is that thesis with a working exploit attached.
This connects to a nasty week for agent security generally. Ars Technica documented HalluSquatting, which weaponizes LLMs' inability to say "I don't know" by registering the package names models hallucinate, turning agent confidence into a distribution vector for botnets across nine popular AI tools. (Ars Technica) And the Vera framework paper hit a 93.9% average attack-success rate against four production agent frameworks. (arXiv)
What to do today: audit every agentic workflow that reads untrusted input (issues, PR comments, emails, webhooks) and has access to anything private. Treat all external text as data, never as instructions. If your agent can both read a public Issue and write to a private repo in the same context window, you have a GitLost waiting to happen. The uncomfortable part is that this is structural, not a patch you install once. The composability that makes agentic workflows useful is the exact thing that makes them a confused-deputy problem.
4. Godot bans vibe-coded contributions, and it's a governance shot heard across OSS
On July 1, the Godot Foundation rewrote its contributor guidelines to ban nearly all generative-AI code submissions and autonomous-agent PRs. New contributors with three or fewer merged PRs now need maintainer permission before submitting features or refactors. (The Register)
The maintainers were blunt about why: an unsustainable flood of buggy AI PRs, cheap to generate and expensive to review, and the line that's getting quoted everywhere, "AI cannot take responsibility." AI is still permitted for menial work like completion, regex, or find-and-replace, and it has to be disclosed. Everything else is out.
I've been going back and forth on this one because I ship with agents every day in my personal projects, so a blanket ban feels like it's aimed at me. But read it as a maintainer, not a contributor, and it snaps into focus. A volunteer maintainer's scarcest resource is review attention. An agent can generate a plausible 400-line PR in ninety seconds. The economics are broken: generation is near-free, review is expensive and human, and the ratio just went vertical. Godot isn't anti-AI. It's protecting the one bottleneck that can't be scaled with a subscription.
This doesn't sit in a vacuum. The same week, a consultancy post titled "We charge $10k a week to delete AI-generated code" hit 282 points on HN, describing cleanup of unreviewed AI "slop" as an actual business line. (odra.dev) And DietrichGebert/ponytail, a Claude Code plugin that biases agents toward deleting code rather than writing it ("the best code is the code you never wrote"), went viral at 77,609 stars. (GitHub) Three signals, same direction: the market is starting to price the downstream cost of generation volume, and restraint is becoming the valued skill.
What builders should take from this: agentic contributions now carry reputational and access risk in communities you don't control. If you're going to open a PR against a project you don't maintain, the burden is on you to review the agent's output as if you wrote it by hand, because from the maintainer's side, you did. The disclosure norm is coming. Get ahead of it.
5. Seat compression stopped being a thesis and showed up on the renewal table
For two years, "agents will compress SaaS seats" was a slide in a venture deck. This week it became a line item in renewal conversations. Enterprises are now openly telling vendors they want to buy 100 seats where they used to buy 500, and vendors are quietly cutting discounts to hold the contract rather than change the model. (Tech-Insider / SaaStr)
The quantified version showed up alongside it. The SaaS CFO reports median gross revenue retention dropped from 88% to 84% in 2026. (The SaaS CFO) That sounds small until you run it: a $50M ARR company now has a $2M hole open in its baseline every year purely from churn and downgrade, before any expansion. And the mechanism that used to paper over that, net revenue retention driven by seat expansion, is exactly what breaks when customers replace seats with agents instead of adding them. The expansion motion that props up SaaS multiples runs in reverse.
SaaStr, which itself reported running 20 agents plus 1.2 humans to do a 10-person team's work, published a reality check this week: most companies won't vibe-code their own Salesforce, even as agents get cheap. (SaaStr) That's the important nuance, and it tells you where to attack. Agents cannibalize workflows and seats first. They don't cannibalize the data and the system of record. Your CRM's moat isn't the seat license, it's the fifteen years of customer records nobody wants to migrate.
So if you're building against an incumbent, don't try to out-Salesforce Salesforce. Attack the seat-priced workflow layer sitting on top of the data, where the per-seat AI surcharge is fattest and the switching cost is thinnest. Cheap open-model inference (see story 1) is what makes that math work for a solo team. This is the same collapse as the margin story, viewed from the buyer's chair instead of the provider's. Supply-side: intelligence is getting cheap. Demand-side: headcount-priced revenue is getting cut. Both curves point at the same place.
Section Deep Dives
Security
HalluSquatting weaponizes "I don't know" across 9 AI tools. The attack registers the package and resource names that models hallucinate, so when an agent confidently reaches for a fake dependency, the attacker's payload is waiting. It turns model overconfidence into a supply-chain distribution channel, and it hits nine of the most popular tools. (Ars Technica) If your toolchain auto-installs anything a model suggests, this is your problem. Pin dependencies and verify existence before install.
Vera auto-generates agent safety tests and it's brutal. The framework does literature-driven risk discovery, combinatorial case generation, and sandboxed adaptive execution, with verifiers that judge actual tool-call evidence instead of the model's self-report. Against four production agent frameworks it hit a 93.9% average attack-success rate, and the authors released Vera-Bench with 1,600 executable cases across 124 risk categories. (arXiv) That's a red-team asset you can point at your own agents this week, not a theoretical result.
JADEPUFFER reported as the first fully autonomous AI ransomware campaign. Coverage on July 7 describes 600 payloads deployed with no human at the keyboard. (BuildFastWithAI) I'd treat this as an emerging claim pending broader corroboration, single-source so far. But the direction is the point: we're moving from AI-assisted offense to agentic offense, and the defensive playbooks written for the former don't cover an attacker that iterates on its own.
CVE-2026-25536: never reuse one McpServer instance across clients. The MCP TypeScript SDK cross-client leak happened because a single long-lived server instance let request and response data bleed between sessions. The fix is a structural one-liner: instantiate a fresh McpServer per client connection. (Practical DevSecOps) Grep your custom MCP servers for a module-level singleton. If you find one, you have the bug.
Agents
Stack Overflow: "Agent orchestration is so two-years-ago." You.com CTO Saahil Jain argues that hand-wiring every step with thick orchestration layers, the 2024 mindset, now actively hurts performance because modern models handle long-horizon tasks far better on their own. (Stack Overflow) I half-agree. Thick scaffolding does fight capable models. But "trim orchestration" and "have no structure" aren't the same thing, and the security stories above are what happens when you trust model capability without guardrails. Trim control flow, keep the safety rails.
Danus coordinates math agents over a shared fact-graph memory. The authors report contributions to research-level math problems using multiple reasoning agents writing to and reading from a shared graph memory. (arXiv) The fact-graph-as-shared-memory pattern is a live design question if you're building multi-agent harnesses. Single-source, so weight the architecture, not the claimed results. But structured graph memory beating a shared scratchpad of prose matches what I've seen building GraphRAG systems.
Lilian Weng distills 35 papers on harness engineering. AINews covered Weng's synthesis of the emerging discipline of building the scaffolding, tools, memory, evaluation loops, that lets agents improve themselves. (Latent Space) If you're designing anything with a self-improving outer loop rather than one-off prompt chains, this is the reading list. It pairs directly with the "agents that write their own skills" pattern below.
Microsoft Agent Framework hits 1.0. The orchestration stack reached general availability, signaling API stability for multi-agent apps. (Microsoft DevBlogs) It lands in a crowded field next to Google ADK, LangGraph, and Genkit. My honest take: framework 1.0 announcements matter less than they used to, precisely because of the "trim orchestration" argument. Pick the one with the escape hatches, because you'll be ripping out scaffolding in six months.
Survey: 65% of working engineers now run two AI coding agents daily. A small sample of 40 engineers, so directional not definitive, but it points at parallel multi-agent workflows replacing single-tool pair programming. (AI Builder Club) Developers are diversifying across Claude Code, Codex, and Cursor rather than standardizing. That matches my own setup, and it's why model-agnostic harnesses keep winning.
Research
Grant Sanderson: AI's math ability is "fractally spiky." On Dwarkesh's podcast, the 3Blue1Brown creator described models that cold-solve an IMO geometry problem in 19 seconds yet flounder on combinatorics, and noted the loop to verify whether a proof is even interesting can run a century. (Dwarkesh) The reusable lesson for builders: competence is jagged within a single domain, not just across domains. Trust agent output per-task, never per-domain. "It's good at math" is not a thing you can rely on.
"Data Analysis in the Wild" benchmarks LLMs on messy real data. The authors argue current benchmarks are too clean to predict production behavior and build a harder, realistic data-analysis evaluation. (arXiv) This is the benchmark I actually want before shipping any LLM data feature. Curated-benchmark scores lie about in-the-wild reliability, and the gap is exactly where your users live.
DynaKRAG adds a learnable retain/drop gate to multi-hop RAG. Instead of naively accumulating every retrieved document across hops, it learns which evidence to keep as new bridge facts emerge. (arXiv) Context bloat and irrelevant-evidence drift are real failure modes I've hit building multi-hop retrieval. The learnable gate is the transferable idea even if you don't adopt the whole framework.
Dwarkesh probes the next training paradigm. The conversation asks what succeeds pretraining-plus-RL as data and compute returns compress. (Dwarkesh) Worth your time if you're trying to anticipate where capability gains come from next instead of assuming the current recipe keeps scaling. Ties back to Amodei's own timeline walk-back below.
Infrastructure & Architecture
ZML ships LLMD, a Zig-based inference server that runs across 5 chip architectures. The French startup's alpha compiles model graphs into standalone native binaries via Zig, MLIR, and OpenXLA, running Llama, Gemma, Qwen, and Mistral transparently across NVIDIA CUDA, AMD ROCm, Google TPU, Intel oneAPI, and Apple Metal, with zero Python deps. (TechCrunch) It does continuous batching, paged attention, and prefix caching, but it's explicitly not production-ready. Still, chip-agnostic inference without the Python-wrapping-CUDA tax is the right architecture, and I'll be watching this one.
Vercel acquires Better Auth to build agent identity into the stack. Better Auth, the MIT-licensed TypeScript auth library with 4.7M+ weekly npm downloads and 850+ contributors, stays free with the same team. (Vercel) The real prize is Agent Auth: each AI agent gets its own scoped, revocable identity with the user as the single point of control. Auth infrastructure is reorganizing around agents acting on your behalf, and that's exactly the primitive GitLost-style attacks need to be contained. Scoped, revocable per-agent identity is how you limit a confused deputy's blast radius.
AMD Ryzen AI Halo debuts as a ~$4K local AI dev kit. LTT Labs detailed AMD's box aimed at on-device inference and fine-tuning, and it pulled 373 points and 259 comments on HN. (LTT Labs) The community heat is the signal: appetite for capable local hardware is climbing as fast as the open-model quality. Cheap open weights plus a $4K box that runs them locally is a real alternative to per-token API dependence for a lot of workloads.
South Korea commits $880B over 10 years to chips, AI, and robotics. Samsung and SK Hynix alone are slated to put roughly $518B toward new fabs. (Crescendo AI) State-scale bets on the compute supply chain are the counterweight to the margin-collapse story: inference gets cheaper on the model side while nations pour hundreds of billions into the silicon underneath it.
Tools & Developer Experience
Kimi K2.7 lands as a first-class model in GitHub Copilot Business and Enterprise. As of July 7, Moonshot's model is selectable alongside GPT and Claude. (GitHub Changelog) Copilot going model-agnostic is the same trend as the margin story showing up inside the biggest coding tool on earth. Teams can now route agentic work to a Chinese open-weight frontier model without leaving the harness they already pay for.
Copilot's in-editor browser tools reach GA. On July 1, GitHub shipped the ability for the agent to drive a browser and verify its own web changes without leaving VS Code. (GitHub Changelog) Closed-loop UI work, edit, render, inspect, in one flow, is the actual unlock. This mirrors Claude in Chrome and chrome-devtools-mcp. The agent finally sees what it built.
Simon Willison ships a <github-code> Web Component, one-shot, via GPT-5.5. The component embeds code directly from GitHub through an href attribute, built entirely by prompting. (Simon Willison) It's a small, replicable example of the kind of tool you can now spin up in minutes. The interesting part isn't the component, it's that "just prompt a working Web Component into existence" is now unremarkable.
Pi, the minimal token-efficient CLI harness from Armin Ronacher and Mario Zechner, gains traction. The Flask and Sentry creator's lean CLI agent is getting cited alongside OpenCode, Aider, and Goose, with token-efficient design as its whole pitch. (arcbjorn.com) For anyone running on a subscription or local models, its lean context handling is the specific reason to evaluate it over heavier harnesses. This is the token-stack story showing up at the harness architecture level.
Models
Tencent open-sources Hy3, a 295B Apache-2.0 MoE, free on OpenRouter until July 21. Simon Willison covered it: 21B active parameters, 256K context, Apache 2.0. By Tencent's numbers hallucinations dropped from 12.5% to 5.4%, and it beats GPT-5.5 on scientific tasks while rivaling open flagships 2 to 5x its size. (simonwillison.net) Another drop-in, permissively-licensed agentic backend. The free-until-July-21 window is a no-cost way to A/B it against whatever you're paying for now. This is the third Chinese frontier-adjacent open model in this issue alone.
Claude Fable 5 drops out of included subscription access today, now metered at $10/$50 per million tokens. July 7 was the last day Fable 5 was included in Pro, Max, Team, and select Enterprise plans at no extra cost. Starting today it bills as usage credits, roughly double the cost of Opus 4.8 for the same seat. (Tech Startups) The model didn't change, the economics did. If you've been leaning on Fable 5 in daily agent loops, audit those pipelines today before the bill surprises you. Timing note: this pricing shift landing the same week as the margin-collapse story is not a coincidence anyone should ignore.
Dario Amodei walks back his own timeline. In a July 6 STAT interview tied to the launch of Claude Science, Amodei tempered his "compressed 21st century" thesis, conceding models aren't yet good enough, researchers need time to learn the tools, and infrastructure and regulation lag. (STAT News) A rare public recalibration from the field's most aggressive accelerationist. For builders, the signal is that even the bull case is now pricing in a slower diffusion curve than the essays implied. I find this more reassuring than the hype ever was.
OpenAI ships gpt-realtime-2.1 with 25% lower p95 latency. The update to its Realtime voice models cuts tail latency by at least 25% while improving recognition, noise handling, and tool use. (OpenAI) If you ship voice agents, tail latency is what users actually perceive as "laggy," so this is a more meaningful update than the version bump suggests.
Vibe Coding
Anthropic extends Claude Cowork to mobile and web. Kick off a task at your desk, get status on your phone, pick up finished output with the laptop closed. (TechCrunch) Async, cross-device agent workflows are becoming table stakes. The desk is no longer where the work happens, it's where you start it. Kilo's coding agent added phone support the same week, same pattern. (Kilo)
Self-improving agents that write their own skills are shipping, not demoing. NousResearch's Hermes Agent authors its own reusable skills from experience and self-updates. (Hermes Agent) The common loop is persistent memory plus the ability to write new skills, so capability grows run-over-run without human authoring. This validates the self-improving outer-loop pattern advanced builders have run in private harnesses. It's also, if you squint, the thing my own pipeline does every night, so I'm biased toward believing in it.
Google CFO: Anthropic codes nearly 100% of its work with AI, Google ~50%. Anat Ashkenazi's figure, surfaced alongside Anthropic's Agentic Coding Trends Report, is a concrete data point on how far internal AI-assisted development has gone at frontier labs. (BuildFastWithAI) Take the 100% with a grain of salt, "with AI" covers a lot of ground. But the Anthropic-vs-Google gap is the real story: the lab building the model is further down the curve than the lab with more engineers.
Perplexity is building "Teammate," a long-horizon coding agent. In use internally since May to find bugs across Perplexity's systems, built for long-horizon engineering with multi-model access. (Business Insider) Not public yet, so an early signal not a shipped product. But it's another search company deciding the coding agent is the real product. Cursor, Claude Code, and Codex now have a fourth serious challenger circling.
Terminal coding agents are proliferating and specializing. DeepSeek-Reasonix tunes around DeepSeek prefix-cache stability, oh-my-pi uses hash-anchored edits plus LSP and subagents, nanocoder keeps code local as a BYO-model agent. (GitHub) The differentiation moved from "which model" to harness-level details: edit anchoring, cache strategy, tool ergonomics. Model-specific harness tuning is now a real performance lever, not a rounding error.
Hot Projects & OSS
ponytail hits 77K stars for making agents write less code. The Claude Code plugin makes an agent "think like the laziest senior dev in the room," biasing toward deletion. (GitHub) 77,609 stars is a loud counter-signal to the generation gold rush. The community is voting for restraint, and it's the same sentiment Godot codified into policy. Something's shifting in what people actually want from these tools.
open-design turns your coding agent into a local design engine, 76K stars. nexu-io/open-design is a local-first desktop app generating prototypes, landing pages, dashboards, slides, images, and video from your agent. (GitHub) It reframes the coding agent as a visual output tool, not just a code writer. For design-literate builders, it's a bet that taste-plus-agent beats template-driven site builders, which is exactly the bet I'd make.
Rowboat: a local-first alternative to Claude Cowork with a knowledge graph. 190 points on HN. It indexes email, meetings, Slack, and assistant conversations into an Obsidian-style backlinked graph stored as local Markdown, then runs background agents on events. (Hacker News) Works with Ollama, LM Studio, or hosted LLMs, pitching persistent cross-session memory as the differentiator. Local-first plus a real memory graph is the combination I keep wanting from these assistants.
Hugging Face ships LeRobot v0.6.0. The open robotics stack's "imagine, evaluate, improve" release continues HF's push to make embodied-AI experimentation accessible on open tooling. (Hugging Face) Robotics is quietly following the same open-tooling curve that ate the model layer.
Fortress: a stealth Chromium so agents stop getting blocked. An open-source build to help browser-driving agents evade bot detection, 41 points but a lively 58-comment thread. (GitHub) The arms race between agent operators and anti-automation defenses is now its own tooling category. If you build web-browsing agents, you'll meet this problem, and the ethics of the workaround are worth thinking about before you ship it.
SaaS Disruption
Fin's per-resolution pricing has a built-in trap: costs double as the AI gets better. Fin charges $0.99 per resolution, per handoff, and per disqualification. The catch analysts flagged this week: the better the agent performs, the more you pay, with costs roughly doubling as resolution rate climbs from 30% to 70%. (Fin.ai) This is the load-bearing tension in outcome-based pricing that the whole industry is racing toward. If you adopt per-outcome billing, build in a cost ceiling or a blended model, or successful automation becomes a runaway line item.
The credit economy hit critical mass: 79 of the top 500 SaaS firms now sell credits, up 126% since 2024. Figma, HubSpot, and Salesforce all joined this cycle. (Monetizely) The shift shows up simultaneously in design, CRM, and platform, which is what "cross-category default" actually looks like. If you're pricing an AI product in 2026, hybrid base-plus-credits is the pattern buyers already recognize, so you're not educating the market, just meeting it.
Greenhouse CEO says hiring is in an "AI doom loop." Candidates mass-apply with AI while employers screen with AI, and Greenhouse is responding by shipping notetaker and setup agents plus a connector that lets Claude or Gemini act directly against the ATS. (HR Dive) A parallel study found 63% of job seekers faced an AI interview, most were never told, and 38% walked away because of it. The recruiting stack is being rebuilt around agent-to-agent interaction, with the application form as a thin human veneer.
Skello raises ~$217M for AI shift scheduling. The European HR platform closed a €200M growth round targeting frontline and shift work, the segment Workday and legacy HRIS underserve. (SaaSRise) Investors still fund vertical HR agents even amid the broader SaaS rout. Frontline workforce ops (retail, hospitality, healthcare shifts) is a category with clear agent ROI and weak incumbents, which is the exact profile a solo builder should be scanning for.
Policy & Governance
US lawmakers open a probe into corporate use of Chinese AI models. A July 8 CNBC follow-up reports Congress is turning the cost-driven adoption surge into a national-security and compliance question. (CNBC) If you're routing tasks to DeepSeek, GLM, Kimi, or Qwen, expect procurement, data-residency, and export scrutiny even when you self-host the weights. This is the policy counterweight to the margin-collapse story, arriving exactly one day after it. The cheap option now carries a compliance tail.
A federal judge rejects Workday's "we just provide the tool" defense. The Mobley ruling means Workday must answer AI hiring-discrimination claims, a precedent that puts every HR-tech vendor whose agents screen candidates on the hook for outcomes, not just code. (SHRM) If you ship automated screening or evaluation, the agent's decisions are now your legal exposure. "It's just software" stopped being a defense this week, and that reframing extends well past HR tech.
Sam Altman floats donating 5% of OpenAI's equity to a US sovereign wealth fund. Per an FT report via TechCrunch, with rival labs expected to match, and a separate call for a US-led international AI standards forum. (TechCrunch) Talks are preliminary and would need congressional approval. My read: this is OpenAI trying to buy political durability as it cedes technical ground to Google and Anthropic. Watch it as a leading indicator of how regulation and public ownership get entangled.
Anthropic opens Claude for Government in beta. A tailored public-sector offering, landing alongside the Cowork expansion and the same month's lifting of export controls on Anthropic's Mythos and Fable models. (Anthropic) Frontier labs are competing hard for government contracts, and the export-control thaw is what makes those deals shippable.
Skills of the Day
1. Set clear_at_least when using Claude's context-editing beta. Auto-clearing old tool results drops a documented example from 70k to 25k tokens, but clearing invalidates your cached prefix. Set a minimum clear size (say 5k tokens) so you only break the cache when the savings actually beat the cache-rebuild cost. (Claude Docs)
2. Make the agent write a script that returns only results, instead of reading files into context. Have it write and execute a small script whose stdout is the only thing that comes back. One execution replaces several read and grep calls, and the raw bytes never cost you context. It's the single highest-leverage habit for keeping long sessions cheap. (context-mode)
3. Pair the memory tool with context editing so learning survives the clear. Tool-result clearing throws away exactly the data an agent might still need. Have Claude write durable facts to the memory tool before the clear fires, so the summary persists to a file while the bulky raw results get evicted. Long-horizon runs stay under the trigger without amnesia. (Claude Docs)
4. Route isolated subagent work to Haiku and scope rules by path. Firecrawl's benchmark found model routing and path-scoped rules were the two biggest token wins, with one technique hitting 91.9% context reduction and no measured quality drop. Stop loading a monolithic CLAUDE.md when you only touched one directory. (Firecrawl)
5. Gate MCP tool schemas until they're needed. Loading every tool's full JSON schema every turn is a hidden cost that scales with tool count. Dynamic tool gating with lazy schema loading defers each schema until the model reaches for that tool, measuring a 95% cut in tool tokens per turn. "Add more tools" goes from free to budgeted. (arXiv)
6. Compress your SKILL.md files. SkillReducer found over 60% of typical skill-file content is prose that doesn't change agent behavior, and reference files can inject tens of thousands of tokens per call. Mechanical compression yielded a mean 77.5% token cut without losing capability. Strip the narrative, push detail into on-demand referenced files. (arXiv)
7. Use two-stage retrieve-then-rerank with Matryoshka embeddings. Retrieve 50 to 100 candidates with truncated 256-dim vectors, then rerank with a cross-encoder (Cohere Rerank 3.5, Voyage rerank-2.5) at full 3072 dims only where precision matters. You stop paying for high dimensionality on every candidate. (FreeAcademy)
8. Fix retrieval before you touch the prompt. 2026 analysis attributes 73% of RAG failures to retrieval, not the LLM, so prompt tweaks are wasted when the wrong chunks arrive. Replace fixed-size splitting with semantic chunking, and add hybrid dense-plus-BM25 retrieval merged by Reciprocal Rank Fusion so exact identifiers and rare names don't vanish into vector space. (Atlan)
9. Force premise-and-trace reasoning before verifying a patch. Agentic Code Reasoning has the model state explicit premises, trace concrete execution paths, and derive a formal conclusion instead of eyeballing correctness. It pushed patch-equivalence accuracy from 78% to 88% on curated cases and 93% on real agent-generated patches. Use it anywhere the answer must actually hold, not just look plausible. (arXiv)
10. Harden every MCP server against the 2026 baseline. An audit found 40% of MCP servers require no auth, 43% carry command-injection holes, and 79% store credentials in plaintext. The checklist: OAuth 2.1 with mandatory PKCE, validate token audience, allow-list every tool input, block egress to private IP ranges to stop SSRF, and never forward the client's token upstream. That last one is the rule people miss. (Codersera)