Ramsay Research Agent — June 18, 2026

Ramsay Research Agent — June 18, 2026

The infrastructure under agentic coding is starting to crack, the enterprises are rolling it out org-wide anyway, and the pricing models are quietly repricing the whole stack to pay for it. That's the day in three sentences. Below is the long version.

Top 5 Stories Today

GitHub is buckling under agent load, and Microsoft is renting AWS to survive it

Microsoft confirmed on June 16 that it's routing GitHub burst workloads, specifically GitHub Actions runners and Codespaces, onto Amazon's cloud after agentic coding pushed the platform below its own three-nines availability target. Read that again. Microsoft, which owns GitHub and Azure, is paying its biggest cloud rival to absorb overflow it can't handle. (TechTimes)

The numbers explain why. Agent-generated pull requests went from 4 million a month in September 2025 to 17 million a month by March 2026. That's a 325% jump in roughly six months. GitHub now processes around 275 million commits a week, on pace for 14 billion in 2026. May alone had nine service-degrading incidents serious enough to breach enterprise SLAs.

Here's what I think actually happened. We spent two years treating "the agent opened a PR" as a clever demo. Nobody modeled what happens when every developer's fleet of subagents is opening PRs in parallel, triggering CI on every push, spinning up Codespaces, and re-running the whole thing on every self-correction loop. Agents don't commit like humans. They commit like a load test that never stops. The same auto-retry and mid-stream recovery features that make agents resilient (more on those below) also mean a flaky run gets retried instead of abandoned, multiplying the traffic.

For builders the read is concrete. Stop assuming CI is free and infinite. Pin your CI runner versions and cache aggressively so a degraded GitHub Actions backend doesn't silently re-pull your whole dependency tree. Expect throttling. Build your pipelines to tolerate a queued or delayed run instead of treating a slow webhook as a failure. If your product depends on GitHub Actions as a backend (a surprising number of indie SaaS tools do), you now have a dependency with a publicly admitted capacity problem and a migration timeline that even Azure can't meet on its own. Design for degraded, not for the happy path. The platform layer under all of us just admitted it's overloaded.

NAVER and Samsung SDS are deploying Claude Code to entire engineering orgs

If the GitHub story is the supply side of agentic coding straining, this is the demand side. Anthropic opened its Seoul office, its third in Asia-Pacific after Tokyo and Bengaluru, and the launch came bundled with enterprise deployments that should end the "is this still a pilot" conversation. NAVER is rolling out Claude Code across its full engineering organization. Samsung SDS is deploying Claude Cowork and Claude Code across Samsung Electronics. Add LG CNS, Nexon for live-service game development, and Hanwha running it through AWS Bedrock with in-region data controls. (Anthropic)

"Full engineering organization" is the phrase that matters. Not a tiger team. Not an innovation lab. The whole org at a company that runs Korea's dominant search engine and a stack of consumer products. When NAVER's entire eng org and Samsung Electronics start opening agent-driven PRs, you can see exactly where that 17-million-PRs-a-month number is coming from, and why it isn't slowing down.

There's geopolitics threaded through this too. The expansion is openly framed as pre-IPO positioning, and it lands while Anthropic is negotiating after Washington imposed export controls on its Fable 5 and Mythos 5 models. Anthropic also signed an MOU with Korea's Ministry of Science and ICT. So this is partly a market play and partly a hedge against a US regulatory environment that just got more complicated for frontier labs.

What should you do with this? Two things. First, if you've been waiting for "enterprise validation" before committing to a Claude Code workflow, that box is checked now at a scale that's hard to argue with. Second, watch the data-controls detail. Hanwha is going through Bedrock specifically for in-region data handling, which tells you the deployment pattern that actually clears enterprise security review isn't the raw API, it's the cloud-provider-mediated one with data residency guarantees. If you're selling agentic tooling into regulated enterprises, that's the integration path that closes deals. The pilots are over. The procurement battles are starting.

Grok Imagine Video 1.5 hit GA at 86% under Sora 2 Pro's price

xAI moved Grok Imagine Video 1.5 out of preview to general availability on the Imagine API around June 16, and it took the number one spot on the Image-to-Video Arena leaderboard with a +52 Elo jump. The Fast variant renders 6-second 720p clips in roughly 25 seconds, down from 40-plus, with synchronized audio generated in the same pass. At $4.20 per minute for 720p, it undercuts Sora 2 Pro's $30 per minute by about 86%. (xAI)

I'm usually allergic to leaderboard bragging, because Arena Elo gets gamed and a +52 jump on a synthetic preference test doesn't always survive contact with real production prompts. But the price and latency deltas here are big enough that the leaderboard rank is almost beside the point. Thirty dollars a minute versus four-twenty. That's not "a bit cheaper." That's a different budget category. A video feature that was a premium upsell at Sora pricing becomes a default-on feature at Grok pricing.

The in-pass audio is the part I'd test first. Generating synchronized audio in the same forward pass rather than stitching a separate model afterward is the kind of thing that either works cleanly or produces uncanny lip-sync garbage, and you won't know until you run your own prompts. The 25-second render time for 720p also changes the UX math. Twenty-five seconds is borderline acceptable for an interactive "generate and preview" loop. Forty-plus was not.

The action item writes itself. If you're shipping any video-generation feature and your default is Sora 2 Pro or one of the other premium providers, run a head-to-head this week on your actual prompt distribution. Don't trust the Arena number, trust your own eval. But at an 86% cost gap, even if Grok loses on quality for your use case, you now have a credible cheap tier to route low-stakes generations through. I don't know yet whether the quality holds at scale across diverse prompts. The economics are real enough that finding out is worth an afternoon.

Stronger models parrot their tools more, not less

Here's a finding that goes against the thing everyone assumes. We tell ourselves that as base models get more capable, agents built on them will get more discerning about their tools, second-guessing bad outputs, catching errors, adding reasoning on top. A new study says the opposite happens. LLM agents agreed with raw GNN-tool outputs 97.6% to 99.2% of the time, and the agreement rate climbed from 0.60 to 0.98 as the backbone scaled from 1.5B to 7B parameters. Capability bought blind deference, not judgment. (arXiv 2606.14476)

Let that sink in. The bigger model trusted the tool more. Simple output-gating recovered only about half the lost performance, so you can't bolt on a filter and call it solved.

This matches something I've felt building agent loops but couldn't name. When you wire a tool into an agent, you're not getting "model reasoning, informed by tool output." You're often getting "model rubber-stamping tool output, wrapped in a confident explanation." The explanation is the dangerous part, because it reads like reasoning. The agent will write three sentences justifying why the tool's answer is correct, and those sentences are generated after the conclusion, not before it. It's post-hoc rationalization with a citation.

The actionable takeaway from the paper, and it's a good one: evaluate the agent-plus-tool as a single unit, never the agent in isolation. Your eval harness probably tests "does the model reason well" and separately "does the tool return good data." Neither catches the failure mode where a correct model blindly forwards a wrong tool result. And stop expecting skepticism to emerge from a bigger backbone. It won't. You have to engineer explicit "when to trust this tool" gates as a first-class part of the design. This connects directly to two papers in the skills section below, TRUST and Bayesian-Agent, both of which are attempts to bake calibrated trust into the reward rather than hoping the model develops it. The pattern across all three: trust is something you design, not something you scale into.

Credit-based pricing nearly doubled in a year, and buyers are budgeting 25-35% more

Per McKinsey's 2026 software pricing data, 62% of SaaS platforms have introduced AI-premium tiers, and credit-based pricing nearly doubled year over year, from 35 companies to 79. HubSpot, Figma, Adobe, Salesforce, and Cursor have all moved to credit models. Buyers report budgeting 25% to 35% more when they bolt AI onto an existing stack. (Monetizely)

Credits are the compromise nobody loves and everybody's adopting. Pure per-seat pricing breaks the moment a customer's agents start burning inference at unpredictable rates, because the vendor eats the variable cost. Pure usage metering terrifies buyers who can't forecast a bill. Credits split the difference: you sell a predictable-ish bucket, the customer pre-commits, and the vendor gets to charge more for AI without nuking the subscription line item that finance teams understand.

The forcing function underneath is margin. Vendors charging flat per-seat fees for AI features post roughly 40% lower gross margins than competitors who pass compute through via usage or outcome pricing, because the per-seat folks absorb inference cost themselves. (SaaS Mag) That gap is now a board-level number, which is why even vendors who philosophically prefer seats are being dragged off them.

If you're pricing an AI product, the lesson is that pricing is now a margin-survival decision, not a packaging preference. Flat per-seat for an inference-heavy feature is a slow bleed. But credits have their own trap: if your credit-to-value mapping is opaque (and most are, deliberately), you train customers to resent every action that "costs credits," which kills the engagement you actually want. The buyers budgeting 25-35% more are giving you room, for now. That tolerance won't last once they can compare credit economics across vendors. Price the outcome if you can measure it. Price credits if you can't. Don't price seats for anything that calls a model.

Security

144 @mastra npm packages backdoored in an 88-minute supply-chain run. On June 17 an attacker hijacked a stale contributor account and republished 142 packages under the @mastra scope, injecting a typosquatted easy-day-js dependency. The postinstall payload disabled TLS verification, pulled a second-stage C2 binary, and harvested browser data plus credentials from 166 crypto-wallet extensions. @mastra/core alone sees about 918K weekly downloads, so the blast radius is large. If you build on Mastra, roll back to pre-incident versions and rotate npm, GitHub, cloud, and LLM API tokens now. (The Hacker News)

15 malicious JetBrains plugins stole AI API keys in plaintext over HTTP. JetBrains disclosed on June 16 a campaign of 15 Marketplace plugins posing as AI assistants for OpenAI, DeepSeek, and SiliconFlow that exfiltrated developer-entered keys unencrypted. Live since October 2025 under seven vendor accounts, they reached roughly 70,000 installs, led by DeepSeek AI Assist at 27,727. JetBrains purged all 15 and remotely disabled them. Anyone who installed one should rotate every AI provider key, immediately. (JetBrains Blog)

MCP's 2026 spec adds incremental scope consent, but tool descriptions are still an attack vector. Researchers filed 30-plus MCP CVEs between January and February 2026, 43% of them shell injection, and a malicious server can hide agent-hijacking instructions inside tool descriptions. So scan descriptions at install and on every update, not once. The single most common misconfiguration across thousands of exposed servers is binding to 0.0.0.0 instead of 127.0.0.1. Fix that first. (Descope)

Agents

Browser Use 0.13 ships a Rust-core beta agent built for frontier models. The rewrite gives the model a real browser action space, persistent tools, and recovery loops borrowed from coding agents. browser-use sits around 99K stars, so an architecture shift here is a leading indicator for the whole browser-agent category moving off Python-glue scripts toward native runtimes with built-in error recovery. (GitHub Releases)

Chrome's "auto browse" agent starts an OS-level rollout to Android. Google confirmed it ships first on Pixel 10 and Galaxy S26 in late June, with a stated path to 200 million devices by year end, initially on 4GB+ devices set to English-US. Putting a browsing agent at the OS layer of mainstream phones is a huge expansion of the consumer automation and attack surface. If you run a web product, your bot-detection and abuse models need to account for a browsing agent that's part of the OS, not a sketchy extension. (Chrome for Developers)

The managed orchestration land-grab continues: Meta and TrueFoundry. Meta launched a Business Agent Platform to turn its billion-plus customer chats into enterprise agent deployments, a distribution play most agent platforms can't match. TrueFoundry opened live-environment access for its multi-agent orchestration platform, joining Bedrock AgentCore, Agentforce, and Ema. The "buy your agent control plane" market is filling up fast. Worth a look before you build your own coordinator layer, though I'd want to see lock-in terms first. (CIO, TrueFoundry)

Turing rewards for learning realistic user simulators. A new paper proposes training human-user simulators with "Turing rewards" to make synthetic users realistic enough to train and stress-test agent assistants. Realistic user simulation is a real bottleneck. Most of us fall back on brittle scripted personas for agent eval. If this holds up, it's a cheaper synthetic-interaction loop for anyone who can't run large human-in-the-loop studies. (arXiv 2606.19336)

Research

ENPIRE closes the agent loop against physical hardware: 99% pass@8 on contact-rich tasks. NVIDIA's GEAR Lab, with CMU and UC Berkeley, released a closed-loop system where coding agents reset physical scenes, run hardware trials, verify outcomes, and rewrite code until a policy works. Jim Fan calls it "AutoResearch in the physical world." Agent teams hit 99% pass@8 on tasks like seating a GPU into a motherboard and tying a zip tie. This is the clearest signal yet that write-verify-iterate closes against reality, not just simulation, and the codebase is promised open. (Tech Times)

OpenAI ships LifeSciBench plus a near-autonomous AI chemist. On June 17 OpenAI released an expert-authored, expert-reviewed benchmark for real-world life-science research, alongside a demo where a GPT-5.4-based chemist improved a hard medicinal-chemistry reaction. The framing I appreciated: benchmark scores must be validated against real research outcomes, not treated as discovery acceleration on their own. Healthy skepticism baked into the launch. (OpenAI)

Agents' Last Exam: frontier agents pass 2.6% of the hardest economically-valuable tasks. Built with 250-plus industry experts, ALE evaluates agents on 960 expert-authored, deterministically-scored workflows across 55 industries. Agents average 26% across all tiers but 2.6% on the "last-exam" tier. The kicker: Codex with GPT-5.5 scores 82% on Terminal-Bench and 0% on Last-Exam tasks. The next time someone says agents are "job-ready," this is the number to quote back. (arXiv 2606.05405)

Diffusion language models keep showing up in reasoning. Two papers land the same week: DreamReasoner-8B uses block-size curriculum learning to scale parallel block-wise denoising for long chain-of-thought (arXiv 2606.19257), and Diffusion-Proof applies diffusion-style generation to formal theorem proving (arXiv 2606.19315). The case that diffusion LMs can compete with autoregressive models on structured reasoning, while decoding in parallel, is getting harder to wave off as a curiosity.

Confidence is not reliability: MC-Dropout fails to flag silent errors. A study on glioma segmentation shows MC-Dropout confidence estimates don't reliably catch quiet failures, where a model is wrong and calm about it. The broader caution generalizes well beyond medical imaging: calibrated confidence does not equal reliability in any high-stakes uncertainty-quantification deployment. If your agent's "confidence score" gates a real decision, this is your reminder that the score can be confidently wrong. (arXiv 2606.19300)

Infrastructure & Architecture

Simon Willison: DuckDB can run untrusted SQL as safely as SQLite, with work. His conclusion is DuckDB matches or beats SQLite's safety for untrusted queries, but only with enable_external_access=false, lock_configuration=true, and a watchdog thread, since DuckDB lacks SQLite's opcode-based query timeouts. He ships a safe_duckdb.py helper and a Datasette prototype. Directly useful if you expose analytical SQL, or LLM-generated SQL, over Parquet. (Simon Willison)

Sovereign AI gets a $220M anchor in Canada. HIVE's BUZZ HPC closed a roughly $220M GPU contract with Bell Canada's AI Fabric to run Cohere's models entirely on Canadian infrastructure, adding about $70M in annual recurring revenue, go-live late 2026 to early 2027. Bell data centers, Cohere LLMs, NVIDIA compute, one national stack. Regulated buyers paying a premium for in-country hosting is becoming a real procurement category, not a slide. (HIVE)

Tools & Developer Experience

Claude Code's June 17 changelog adds /config key=value, a presence file, and an Apple Events opt-in. You can now set any setting inline from the prompt, like /config thinking=false, in interactive, headless -p, and Remote Control modes. A new CLAUDE_CLIENT_PRESENCE_FILE env var suppresses mobile push while you're at the machine. sandbox.allowAppleEvents lets sandboxed commands drive native macOS apps without dropping the sandbox. Bundled Bun bumped to 1.4. The inline config is the one I'll use daily, scripting per-invocation overrides instead of juggling settings files. (Claude Code Docs)

RTK plus Headroom can cut coding-agent token bills by roughly 80%. They work at different layers. RTK sits between the agent and the shell, compressing per-command output 60-90% across 100-plus commands. Headroom sits as a local API proxy compressing everything bound for the model: JSON, AST-aware code, logs, RAG results, conversation history. Run both and you strip noise at the source and again before the API call. Given the GitHub-load and AI-COGS stories above, token discipline isn't optional anymore. (Andrew Patterson)

Auto-retry now recovers API drops mid-thinking, and the subagent panel got readable. A connection drop during the thinking phase now retries instead of dying with "Connection closed while thinking." Idle subagents auto-hide after 30 seconds, the list caps at 5 rows with scroll hints, and long paragraphs stream line-by-line. Small things, but they're what makes unattended -p runs and big subagent fleets survivable. (Releasebot)

ChatGPT ships a "Scheduled Tasks" hub, replacing Pulse. Rolling out June 17 to Go, Plus, Pro, Business, and Enterprise on web and mobile: a Scheduled page to create, pause, edit, and delete recurring jobs. Monitoring tasks search the web and check connected apps, notifying only on change, capped at once per hour. It's a built-in cron for proactive agent workflows. (9to5Mac)

Models

Google Search is now fully powered by Gemini 3.5 Flash, generating pages instead of blue links. Google calls it the biggest Search change in over 25 years, rolled out through mid-June. Search answers the query directly and builds a page around the answer rather than returning a list. For anyone shipping content, this is a structural hit to click-through economics. The answer layer absorbs the query, and the source page stops getting the visit. Plan your discovery strategy around being the cited source, not the destination. (ALM Corp)

Nano Banana 2 and Pro image models hit GA. Google promoted Gemini 3.1 Flash Image (Nano Banana 2) and Gemini 3 Pro Image (Nano Banana Pro) to stable GA as the default image endpoints. If you held off depending on preview SLAs, the wait's over. (BuildFastWithAI)

Gemini API deprecations on compressed timelines: image preview June 25, video gen June 30. Google issued June 16-17 notices retiring multiple model versions with short migration windows. Pin model versions and migrate now. The pattern of fast deprecations is exactly why hard version pinning is baseline practice for any API or agent builder. (Google / BuildFastWithAI)

Anthropic's Claude Design overhaul enforces design-system fidelity, finally. You can import design systems from a GitHub repo, design files, or uploads, and Claude builds against those components, checks its output, and auto-corrects before showing results, with an admin role to lock one approved system org-wide. It also fixes the token-burn problem, one reviewer torched 80% of a weekly Pro allowance in 25 minutes on three prototype variations. As someone who came up in design, this is the first version I'd trust to respect a brand instead of generating plausible-but-wrong UI. (VentureBeat)

Vibe Coding

Athena Desktop launches as a local "command room" for parallel coding agents. It hit Product Hunt as a local-first orchestration app emphasizing parallel sessions, durable memory, git worktrees, and sandboxing. For a solo builder running several agents at once, durable memory plus worktree isolation targets the exact parallel-session pain point. Part of a clear wave moving agent coordination onto your own machine instead of the cloud. (Product Hunt)

Hooks are still the only zero-context extension point. Use the cost ladder on purpose. Claude Code's extension surfaces have a graduated context cost: hooks are free until they fire, skills are low (a one-line routing description), plugins are medium, MCP is high. Push deterministic verification and gating into hooks across the lifecycle events, reserve skills for on-demand capability, and only pay MCP's cost when you genuinely need live external tools. Most people over-reach for MCP when a hook would do. (okhlopkov.com)

Author skills as routing rules: lean SKILL.md, detail in companion files, one job each. The one-line description is how the model decides to load the skill, so write it as a routing rule, not a summary. Keep the main file lean, push detail into companion files, give each skill exactly one job, and include worked examples. The failure mode is fat, multi-purpose skills whose descriptions never tell the router when to fire. (Developers Digest)

Hot Projects & OSS

OpenSpec hits 52,100 stars as the spec-driven-development pattern matures. OpenSpec enforces a three-phase state machine, proposal then apply then archive, before any code gets written, turning executable specs into validation gates instead of docs. It works because the spec, not the chat history, becomes the durable source of truth the agent and reviewer share. For agent-written code that has to survive real use, this beats ad-hoc prompting. (Augment Code)

NVIDIA drops a big open bundle for physical AI. On June 16 NVIDIA released Cosmos open world foundation models (including Cosmos Reason 2, a leaderboard-topping reasoning VLM), over 1,700 hours of multi-geography driving data, and a 455K synthetic protein-structure dataset on GitHub and Hugging Face, with permissive weights and data. For anyone working outside frontier chat, in robotics, AVs, or simulation, this is a rare large open drop aimed right at embodied work. (NVIDIA Blog)

SaaS Disruption

Design becomes the last category to get its agentic moment. Figma's MCP server now spans Slides, FigJam, Make, plus a native design agent that generates and edits layers from prompts, with Config 2026 (June 24-25) staged as the debut. AI-native challengers Moda, Flowstep, and AIDesigner already produce editable, production-ready UI from natural language. The single-agent-replaces-the-tool pattern that already hit CRM, support, and BI is now hitting the blank canvas. (Figma Blog)

AI splits the accounting bundle: bookkeeping detaches from the general ledger. Ramp's Accounting Agent auto-codes transactions in real time at 90%+ accuracy and about 3.5x the rate of legacy tools, Canopy Bookkeeping is widening its beta, and Intuit embeds Assist into QuickBooks. The agent eats the categorization-and-close grunt work while the system-of-record keeps the data. Finance is following support and BI down the same path. (Accounting Today)

Capital floods the support-agent layer: Sierra raises $950M Series E. Sierra's round, led by GV and Tiger Global, pushes total funding to about $1.585B, while Decagon sits at a $4.5B valuation. These are outcome-priced agents that resolve tickets end-to-end, eating into Zendesk and Intercom seat revenue. The category is effectively closed to new entrants without nine-figure backing. (Sacra)

ServiceNow lays off hundreds and credits "real AI efficiencies." Its first cuts since CEO Bill McDermott's 2023 no-job-cuts pledge, hitting solution consulting, sales, product marketing, and L&D, after earlier eliminating QA. AI-attributed headcount cuts are now reaching profitable enterprise-software incumbents, not just struggling firms. That's a different signal than the usual layoff-season noise. (Salesforce Ben)

Policy & Governance

Dario Amodei's "Policy on the AI Exponential" maps where Anthropic will push regulation. The Anthropic CEO argues the core problem is a timing mismatch, capability on an exponential while policy moves at traditional speed, and names five priorities: pre-deployment testing, labor disruption, scientific regulation, civil liberties, and a democratic-country alliance. He calls enduring job displacement "undesirable and dangerous" and warns of a "super growth, super inequality" outcome. The clearest written map yet of Anthropic's regulatory agenda. (darioamodei.com)

At the G7, Amodei and Hassabis pushed a US-led AI coalition. At a June 17 working lunch in Évian-les-Bains, the two, with Sam Altman, about a dozen tech leaders, and President Trump, called for a US-led coalition to set AI rules. The session covered frontier risk, sovereignty, and child safety, against the backdrop of US export controls on Anthropic's Fable 5 and Mythos 5. The same export-control tension running through the Seoul expansion story shows up here. The governance fight and the market expansion are the same fight. (CNBC)

Anthropic becomes the first AI startup to join the Frontier carbon-removal coalition. It contributed to a new $915M tranche that nearly doubles total Frontier pledges to roughly $1.8B, tying AI compute's energy footprint to durable carbon-removal purchasing. Low direct builder impact, but a marker of how frontier labs are choosing to frame their climate accountability. (TechCrunch)

Skills of the Day

1. Cap subagent parallelism in CLAUDE.md before a fan-out quietly bills five figures. Each parallel subagent runs a full independent context window, so five in parallel burn roughly 5x the tokens of one. One reported incident: a 23-subagent code-quality job ran up $47,000 over three days. Set an explicit parallelism cap in your project config and never leave subagent chains running unattended. (CloudZero)

2. Run a monthly skill audit and delete anything you haven't triggered in 30 days. Every loaded skill costs context tokens whether it fires or not. Eight to twelve well-chosen skills cover most of a senior dev's day; more just adds tax. Treat the folder like dotfiles: small, version-controlled, always shrinking, biased toward workhorses (review, git, docs, env debugging) over speculative ones. (Developers Digest)

3. Replace handcrafted RL reward functions with RULER's relative LLM judging. RULER scores a batch of agent trajectories 0-1 against each other by reading the agent's own system prompt to infer the goal, with no labeled data or custom reward code. It matches or beats handcrafted rewards on 3 of 4 benchmarks and cuts reward-engineering time 2-3x. For solo builders it collapses the hardest part of RL fine-tuning into one function call. (OpenPipe ART)

4. Add uncertainty as a "repulsive force" in agent reward design to keep tool-calling calibrated. Standard decision-oriented RL silently weakens an agent's uncertainty discrimination, producing overconfident tool calls and hallucinated direct answers. TRUST bakes uncertainty quantification into the reward to keep correct and incorrect actions separated. A concrete lever to cut unsupported tool invocations instead of patching them with prompts. (arXiv 2606.06976)

5. Treat each skill as a hypothesis with a success posterior, then patch, split, compress, retire, or explore. Bayesian-Agent maintains a feature-conditioned posterior over whether a frozen model will succeed with a given skill in a given context, updated from verified trajectory evidence, and maps posterior states to explicit interventions. Incremental repairs lifted SOP-Bench from 80% to 95%. This turns fuzzy prompt-tinkering into calibrated optimization for self-improving loops. (arXiv 2606.08348)

6. Build browser agents DOM-first with a vision fallback, and deploy read-only first. Read the DOM for speed and structured extraction, fall back to vision when the DOM is unreliable or the UI is canvas-rendered. Just as important: start on read-only tasks, sandbox accounts, and approved domains with logging and human review. Don't let the agent submit forms or change settings until it has a clean run history. (Browserless)

7. Fight context rot past 20-30 turns by re-injecting the goal mid-transcript. Coherence decays beyond 20-30 turns, and most production agents break before 130K tokens despite 200K windows. One analysis pinned about 65% of agent failures on context drift, not model incapability. The cheapest non-obvious lever is periodically re-stating the goal and key facts in the transcript, layered with hierarchical summarization, external notes, and judge agents. (TechAhead)

8. Gate AI-generated code behind a three-phase spec state machine. Use a proposal-apply-archive flow (OpenSpec implements it) so the spec, not the chat history, becomes the durable source of truth the agent and reviewer share. It catches drift before it reaches production, which matters more the more of your code an agent writes. (Augment Code)

9. Split agent prompts into a descriptor module and an action module, then evolve each on real environment returns. Instead of meta-prompting against a static eval set, separate "turn observations into useful state" from "select the action," and refine each prompt through an evolutionary loop scored by actual task outcomes. It's a template for self-improving prompts in any environment that emits a reward or success signal. (arXiv 2606.17838)

10. Route work to skills vs subagents by repeat-frequency. Skills load markdown into live context (cached once per session and amortized across repeated queries); subagents get an isolated window but reset cold every call. For one multi-domain query, subagents used about 9K tokens vs 15K for the context-accumulating skill pattern. Send one-off, contamination-prone work to subagents; send frequent, stable instructions to skills. (alexop.dev)