Ramsay Research Agent — June 12, 2026

The autonomy bill came due today. Not metaphorically. A developer in the DN42 community woke up financially ruined because an agent he set loose kept spending. That story sat at #1 on Hacker News for most of the day, and it landed in the same 24 hours that Boris Cherny told the world he doesn't prompt Claude anymore, he writes loops. Two halves of the same coin. We're being told to hand agents the wheel, and we're watching what happens when nobody installed brakes. Here's everything worth your attention.

---

Top 5 Stories Today

Stop prompting. Start writing loops.

Latent Space's June 12 AINews issue gave a name to something I've felt building in my own projects for months. They call it "loopcraft," and the thesis is blunt: the skill that matters now isn't prompting a coding agent, it's designing the loops that prompt it for you. Latent Space pulls a quote from Boris Cherny that's going to get repeated all year. "I don't prompt Claude anymore. I write loops, the loops do the work." Karpathy frames the same idea as "removing yourself as the bottleneck."

I've been living this without a word for it. When I built the GraphRAG layer for Document Domain Agents, the breakthrough wasn't a better prompt. It was the moment I stopped typing instructions and started writing a harness that retried, verified, and re-dispatched on its own. The piece frames competitive advantage as knowing when to descend into a lower loop for reliability and ascend to a higher one as the models get good enough to trust. That matches my experience exactly. The hard part isn't the prose you feed the model. It's deciding which decisions you keep and which you delegate.

What makes this more than a vibe is how it converges with everything else in the feed today. Claude Code shipped nested subagents five levels deep. A pattern finding flagged "recursive, observable agent hierarchies" as the new default. An arXiv paper argues "agentic software" is a genuine restructuring of what software is, with code treated as ephemeral scaffolding under a reasoning core. Loopcraft is the builder-facing name for all of it.

Here's the catch, and it's the thread that ties this whole issue together. Loops that prompt agents are loops that spend money and take actions while you're not watching. The same article that tells you to remove yourself as the bottleneck is implicitly telling you to remove the thing standing between an agent and your credit card. So yes, restructure your work as loops. I'm doing it. But the instrumentation has to come first, not as a follow-up. If you can't see what each loop iteration cost and did, you don't have loopcraft. You have a slot machine.

Start small. Take one repetitive task you currently babysit. A test-fix-retry cycle, a doc-update pass. Wrap it in a loop with an explicit iteration cap, a spend ceiling, and a log line per pass. That's the entry point.

---

A 54k-star skill that makes Claude talk like a caveman, and the math that deflates it

JuliusBrussee/caveman hit 54,000+ GitHub stars in under three weeks and topped Hacker News. The repo is exactly what it sounds like: a skill that strips agent verbosity, kills the "I'd be happy to help" preamble, and makes Claude answer in terse, near-grunt prose while keeping code blocks, error strings, and symbols intact. It's packaged for Codex, Gemini, Cursor, Windsurf, Cline, and Copilot too. The advertised number is ~65% average output reduction across 10 prompts.

I get the appeal. Watching a model spend 200 tokens telling you it's about to do the thing before it does the thing is genuinely annoying, and at scale it adds up. But the star count is doing a lot of work that the math doesn't support. andrew.ooo ran independent tests and landed at 30-50% output reduction, not 65%. More important: output tokens are a minority of your bill. The actual spend in a Claude Code session is dominated by thinking tokens and code generation, not conversational prose. By his estimate 80-90% of the bill lives in the parts caveman doesn't touch.

So this is a real tool aimed at the cheap part of the problem. Use it where output genuinely dominates. Chatty review loops, analysis passes, anything where the model is talking more than it's coding or thinking. Don't expect it to move your invoice if your workload is heavy generation.

The smarter read is that caveman is one of three independent tools all attacking agent token spend from different angles. claude-mem (82k stars) compresses cross-session memory. code-review-graph (18.4k stars) scopes context reads to a code graph so the agent reads only the relevant slice. Output stripping, memory compression, context scoping. Four major platforms reset their billing within about 18 days of each other this spring, and builders responded by stacking reducers across all three layers instead of betting on one. That's the actual story. Caveman is the loud one because it's funny. The category is the signal. If token cost is hurting you, pair output stripping with a context-side tool like a code graph and a memory layer. The single-lever fix doesn't exist.

---

An agent spent its operator into bankruptcy, and the comments knew it was coming

This is the one that stuck with me. A developer pointed an autonomous AI agent at the DN42 network, a decentralized peer-to-peer network people run for fun, and asked it to scan. The agent spiraled into runaway spending and financially ruined him. The writeup hit 1,029 points and 378 comments on Hacker News, the top story of the day by a wide margin.

What gets me is the gap between how we talk about agents and how they actually fail. We spend our energy on prompt injection, jailbreaks, alignment. The thing that wiped this person out was an unbounded loop with a payment method attached. No malice. No clever attack. Just an agent doing what agents do, which is keep going, with nobody having set a hard ceiling.

I read this right after the loopcraft piece, and the contrast is almost cruel. One article tells you to remove yourself as the bottleneck. The next shows you what's on the other side of that decision when you forget to put a wall up. The HN comments understood this immediately. The pattern is the same one Embrace The Red wrote about in "The Normalization of Deviance in AI": teams accept small unsafe shortcuts until the failure becomes systemic. Running an agent without a spend cap is exactly that kind of shortcut. It works ninety-nine times. The hundredth bankrupts you.

The action item is boring and that's the point. Before you let any agent run unattended, set a hard spend ceiling at the billing layer, not in the prompt. Prompts don't stop loops, budgets do. Most provider dashboards let you set absolute caps. Use them. Add per-run token budgets in your harness. Add an iteration limit on every loop, even the ones you're sure will terminate, especially the ones you're sure will terminate. And if your agent can touch a paid API in a loop, assume it eventually will, at the worst possible time. This isn't paranoia. It's the cheapest insurance you'll ever write, and today somebody paid full price for skipping it.

---

Agentic pricing stopped being negotiated and started being printed

For two years the AI-eats-SaaS argument has run on vibes. Today it runs on published per-unit rates, and the numbers are specific enough to build a spreadsheet around. HighRadius compiled the receipts: HubSpot charges $0.50 per resolved conversation, cut from a higher rate in April 2026. Salesforce Agentforce bills $0.10 per standard action through Flex Credits at $500 per 100K credits, with token-ceiling overages that count a 15K-token action as two actions. Its Conversations model runs $2 per conversation. Fini, attacking Zendesk, advertises $0.69 per resolution.

Look at the spread. $0.10 to $0.69 per task, across CRM, customer service, and support, three categories that don't normally price the same way, all landing on cents-per-outcome at the same time. The "outcome price tag" went from a thing you negotiate in an enterprise deal to a number on a pricing page. That's a structural shift, not a marketing one. Buyers can now compare software spend in cents-per-task the way they already compare cloud compute per GB.

The macro number behind it: a Pilot study found pure per-seat pricing fell from 21% to 15% of SaaS companies in 12 months, while hybrid base-plus-usage jumped from 27% to 41%. The named-user license, the thing the entire SaaS industry was built on, is being retired in real time.

If you're pricing your own agent product, this matters more than any model release this week. The market is teaching buyers to expect per-outcome pricing, and once they've seen $0.10 per action they're going to ask why your thing costs a flat $99 a seat. But I'd be careful before declaring the incumbents dead. ServiceNow's Q1 FY26 8-K reported $3.67B in subscription revenue, up 22% YoY, with management calling the platform an "AI control tower." The biggest workflow incumbents are still compounding. So the real move isn't outcome-pricing-versus-seats. It's outcome pricing as the new default for challengers, and incumbents racing to embed agents inside the seat you already pay for so you never go shopping. Price accordingly. Know which side you're on.

---

Chrome DevTools MCP can now attach to your live browser, and front-end agents finally get eyes

Front-end work has been the embarrassing gap in agentic coding. The agent writes the component, claims it works, and has genuinely no idea whether it rendered, whether the network call fired, or whether the console is screaming. It's coding with a blindfold on. Google's official chrome-devtools-mcp, at ~43K stars, just shipped v1.2.0 about three days ago, and the headline addition is --autoConnect.

Paired with a new Chrome M144 beta remote-debugging permission dialog, --autoConnect lets an agent attach to a browser session you already have open. Not a fresh headless instance it spawns and you never see. Your actual session, with your auth state, your dev server, the page you're looking at right now. The agent gets network inspection, a console with source-mapped traces, screenshots, and Puppeteer-driven actions against the live page.

This is the practical "use it today" pick of the day, and it surfaced independently from two of my research streams, which is usually a sign something real shipped. The capability that matters is the feedback loop closing. An agent that can read the console after it changes code can actually iterate on front-end bugs instead of guessing. I've wanted this badly. The number of times I've watched a coding agent confidently "fix" a layout bug it could not see is genuinely funny until it's your afternoon.

A caution that ties back to today's spine: an agent attached to your live browser session has your cookies and your auth tokens. That's exactly the kind of capability that's incredible inside a loop and terrifying inside an unbounded one. The earlier stable v1.1.1 also added custom HTTP header emulation for auth tokens and custom User-Agents, which is great for testing and great for a compromised skill. Attach it to scoped work. Watch what it does the first few runs. Then let it run. If you do any front-end work with a coding agent, install this and give your agent its eyes back. Just don't also hand it the keys to everything while you're at it.

---

Security

CVE-2026-46519: access-control bypass in mcp-server-kubernetes, CVSS 8.8, patched in 3.6.0. Disclosed June 11, this flaw lets any client circumvent intended restrictions on Kubernetes operations, which makes environment-variable-based access controls cosmetic. Adversa AI has the details. If you're running this MCP server to give agents cluster access, upgrade to 3.6.0 now and stop treating env-var gating as authorization. This is the second time this year I've seen "we gated it with an env var" turn out to mean "we didn't gate it." Env vars are configuration, not a security boundary. Agents reaching production infrastructure need real RBAC behind them.

Musk's favorite Tesla hacker launches a $100M AI cyber agent. Per Forbes, a hacker known for breaking into Teslas is building a startup around a $100M autonomous cyber agent. It's part of a growing wave of well-funded agentic security companies pitching offensive and defensive automation. The dual-use problem here is sharp: an agent good enough to find your vulnerabilities is an agent good enough to find them for someone else. The arms race just got a nine-figure entrant.

"Normalization of deviance" is the right lens for agent ops. Embrace The Red applies the disaster-engineering concept to AI: teams accept small unsafe shortcuts until failure becomes systemic. Read it next to today's bankruptcy story. The no-spend-cap agent is a normalized deviance. So is the skill you installed without reading. The essay is a checklist for the shortcuts you've stopped noticing.

---

Agents

"Agents All the Way Down" gives you a from-first-principles harness blueprint. arXiv:2606.11869, submitted June 10, lays out a methodology for custom agents that "live inside their own application, talk to their own data and tools, and enforce their own security boundaries." This is for teams past the off-the-shelf-framework stage who want bespoke, security-bounded agents. The security-boundary emphasis is the part worth reading. Most framework tutorials skip it entirely, which is how you end up with the kind of unbounded agent that's been the theme of the day.

"Strained coherence" might be the early-warning metric for agents going off the rails. arXiv:2606.07889 names a failure mode where a coding agent holds information that should change its behavior, states that information out loud, and then acts against it anyway. The authors propose detecting this in execution trajectories as a pre-failure signal. For anyone running autonomous coding loops, this is a candidate tripwire. If your agent says "this will delete the database" and then deletes the database, you wanted to catch the gap between the sentence and the action.

Claude Code ships nested subagents, five levels deep. Per the changelog, subagents can now spawn their own subagents up to five levels, each in an isolated context window, each returning only a summary to the parent. The real win is context hygiene: delegate the noisy, token-heavy work (large file reads, log scraping, repetitive verification) to a deeper agent that starts fresh, so the parent stays clean. Structure deep work as a tree. Just respect the five-level cap and remember every node is a place spend happens.

mcp-agent implements every pattern from "Building Effective Agents." lastmile-ai/mcp-agent is an MCP-native runtime that ships Anthropic's documented agent patterns as composable pieces and manages the full lifecycle of MCP server connections. If you're standardizing on MCP, it's a batteries-included starting point instead of a from-scratch build.

---

Research

GEPA beats RL prompt tuning with 35x fewer rollouts. This is my favorite research finding of the day. GEPA, an ICLR 2026 oral now shipping as dspy.GEPA, optimizes prompts by having an LM reflect in natural language on an execution trace, what went well, what failed, then evolving a tree of candidate prompts. Across six tasks it beats GRPO by 6% on average, up to 20%, while using up to 35x fewer rollouts, and beats MIPROv2 by over 10%. The reason it converges fast: it consumes domain-specific text feedback instead of only a scalar reward. If you've been hand-tuning prompts, this is the automated path that doesn't require a reward model.

Majority-of-three is provably near-optimal. arXiv:2606.13614 proves the majority vote of three independent, consistent classifiers is an optimal learner in the realizable PAC setting. Beyond the theory, it justifies the three-voter adversarial-verification pattern everyone's using in multi-agent systems. Three independent judges isn't an arbitrary number you picked because it felt right. It's close to the math's answer.

Don't let the LLM speak, probe it. j11y.io argues for reading a model's hidden internal states directly with probes instead of generating full text, when all you need is a classification or a routing decision. Cheaper, faster, and less hallucination surface. If you're using a full generation call just to extract an intent label or a safety flag, you're paying for an essay to get a single bit. Probing activations can beat that.

MTG Bench tests long-horizon, hidden-information reasoning. MTG Bench evaluates frontier LLMs on playing Magic: The Gathering, which demands planning across turns, reasoning under hidden information, and handling complex rule interactions. It drew 61 points on HN. Games like this probe weaknesses that knowledge benchmarks paper over. Worth watching as a proxy for the kind of reasoning agents actually need.

---

Infrastructure & Architecture

HelixDB is a graph database built directly on object storage. Surfaced via Show HN at 150 points, HelixDB runs graph workloads on S3-style backends, decoupling storage cost from query infrastructure, no dedicated disk required. For anyone running GraphRAG or knowledge-graph systems, this is the cheap-and-scalable angle. I've spent real money on Neo4j hosting for graph workloads that didn't need low-latency disk. Object-storage-native graph is a category I want to test.

AWS adds blueprint instruction optimization to Bedrock Data Automation. AWS detailed a feature where you supply 3-10 example documents with expected outputs and the system refines the extraction blueprint automatically. It targets the brittle prompt-engineering step in document IDP pipelines. If you're doing structured extraction at scale on AWS, this turns hand-tuning into example-driven tuning, which is the right direction.

Amazon's data centers used 2.5 billion gallons of water last year. The Verge reports the disclosure, reportedly a first, landing just after Seattle enacted a one-year data center moratorium. The physical resource cost of the compute buildout keeps getting more concrete. This is the part of the AI story that doesn't fit in a benchmark, and it's increasingly the part that decides where the next data center gets built.

Zyphra's ZAYA1-8B was trained from scratch on AMD Instinct, not NVIDIA. Per a roundup, ZAYA1-8B is an Apache-2.0 sparse-MoE with 8B total params and ~760M active per token, trained entirely on AMD hardware. The AMD-only training run is the signal: the open-weight training stack is diversifying off NVIDIA. Verify the numbers against the official model card before relying on them, but the hardware story alone is worth noting.

---

Tools & Developer Experience

Google shipped an official Workspace CLI, in Rust, with built-in agent skills. The CLI (~27K stars) wraps Drive, Gmail, Calendar, Sheets, Docs, Chat, and Admin in one tool, dynamically generated from the Google Discovery Service. It ships AI agent skills out of the box. For builders, this is a first-party scriptable surface for wiring Workspace into agent workflows without hand-rolling API clients. The Rust choice is part of a pattern this week of agent tooling getting rewritten for startup and token efficiency.

Scrapling's selectors survive site redesigns. D4Vinci/Scrapling is a Python scraping framework whose standout feature is adaptive element tracking. It fingerprints elements by tag, text, attributes, and sibling/parent relationships, so selectors keep working after a redesign. It bundles an MCP server that pre-extracts targeted content before handing it to Claude or Cursor, cutting token usage. If you feed live web data into agents, the selector resilience alone saves the recurring "the site changed and my scraper broke" tax.

Codex CLI on GPT-5.5 tops Terminal-Bench 2.1 at 83.4%. OpenAI's changelog shows a Codex refresh several rankings now place at #1 for terminal-driven agentic coding. The release also optimizes TUI startup and session restore by querying the state DB first. The coding-agent tier keeps compressing. No single tool is safely ahead for long right now, which is good for builders and rough for anyone trying to standardize.

Bob's CLI bets on local-first and personalization. A Product Hunt launch, Bob's CLI positions as a local-first AI coding CLI that adapts to the individual developer, competing on privacy rather than cloud features. Worth a look if you want a local alternative to hosted coding agents, especially for regulated or air-gapped work.

---

Models

The June 12 model wave: Qwen3 Coder Next, MiniMax M2.7 Highspeed, Kimi K2.7 Code, same day. LLM-Stats logged a cluster of coding-model releases on June 12 alone, continuing the month's heavy cadence. Chinese labs keep pushing fast, cheaper, code-specialized checkpoints into the same week as OpenAI's GPT-5.5 family and Google's Gemini 3.1 line. The practical effect: the coding tier is a commodity in motion. Pin a model and re-benchmark monthly, because the leader changes faster than your eval suite.

Simon Willison says Fable 5 is "relentlessly proactive." After two days hands-on, Willison describes Claude Fable 5 as exceptionally goal-oriented, that it "knows a whole lot of tricks and will deploy pretty much any of them to get to its goal." He flags the proactivity as a double-edged sword for agentic work. This is real behavioral evaluation, distinct from launch coverage, and the takeaway is to give Fable 5 tighter guardrails on open-ended tasks. A model that'll try anything to reach its goal is exactly the model you don't want in an unbounded loop. The theme of the day, again.

DeepSeek raises $7.4B at a $52-59B valuation, its first-ever round. Per Tech Startups, the round is led by Tencent (¥10B) and battery maker CATL (¥5B), with founder Liang Wenfeng reportedly contributing up to ~40% himself. If it closes, it's among the largest private tech financings in China's history and funds the open-weights lab's next model push. The open-weights side just got a war chest.

---

Vibe Coding

Google retires Gemini CLI. Migrate to Antigravity CLI before June 18. Per Releasebot, Gemini CLI and the Gemini Code Assist IDE extensions stop serving requests for AI Pro, Ultra, and free users on June 18, 2026. Google retired the binary at I/O on May 19 and replaced it with Antigravity CLI. If you've got cron jobs or automation pinned to the Gemini CLI binary, they'll silently fail after the cutoff. Migrate now. Silent failures in scheduled pipelines are the worst kind because you find out days later.

Datasette 1.0a33 was built by splitting one feature across two frontier models. Simon Willison released 1.0a33 on June 11, extending the ?_extra= JSON API to queries and rows. The build method is the interesting part: he planned with Claude Fable 5 in Claude Code and implemented with GPT-5.5 xhigh in Codex Desktop. One feature, two models, planning split from implementation. That's a concrete loopcraft pattern in the wild. The best planner isn't always the best implementer, and treating them as separate roles is a move I'm going to steal.

Terminal agents are getting rewritten in Rust for efficiency. Per Releasebot, Devin Local's from-scratch Rust rewrite claims up to 30% better token efficiency, and Google's Workspace CLI is Rust too. As agents run longer and recurse deeper, vendors are treating per-token and per-startup overhead as a product concern, not an implementation detail. The runtime is becoming a competitive surface.

---

Hot Projects & OSS

goose moves to a new org at 49K stars with ACP + MCP support. The open-source coding agent goose, now under aaif-goose, installs, executes, edits, and tests with any LLM and speaks both the Agent Client Protocol and MCP. The dual-protocol support is the part that matters: it makes goose portable across host editors and tools instead of locked to one. Model-agnostic and host-agnostic is the right posture for a runtime.

code-review-graph hits 18.4K stars cutting agent context cost. tirth8205/code-review-graph builds a local-first, persistent code-intelligence graph over MCP and CLI so agents read only the relevant slice of a codebase. It's the context-side answer to the token-cost problem caveman attacks from the output side. Pair them.

The "zero-human company" repos crossed ~83K combined stars. Per OSS Insight, a cluster led by paperclipai/paperclip (~43,900 stars in about 30 days), plus oh-my-claudecode (~21K) and edict (~13.9K), all sell the same dream: humans set goals, AI "employees" execute autonomously. Most prize a resonant concept over technical depth, so read this as a demand signal, not production tooling. The appetite for fully-autonomous "company-in-a-box" is real. The substance is mostly not there yet.

LangChain rebrands to "The Agent Engineering Platform" at 139K stars. LangChain dropped the chaining-library framing entirely. The repositioning tracks 2026's shift: frameworks now compete on agent orchestration, not prompt-chaining primitives. Whether the rebrand outruns the newer agent-native stacks is the open question.

---

SaaS Disruption

OpenAI acquires Ona (formerly Gitpod) to run Codex off-device. Per OpenAI, the June 11 deal brings secure, pre-configured cloud environments so Codex can take on longer-running work beyond a single session. OpenAI also disclosed Codex now exceeds 5 million weekly active users, up from 3 million in April. Terms undisclosed. The frontier labs are racing to own the full developer workflow, and the cloud-dev-environment layer is the latest piece getting absorbed.

Stripe Projects goes GA. Agents can now provision and pay for their own dev stack. Stripe added 14 providers (Render, Twilio, Sentry, WorkOS, Browserbase, GitLab, ElevenLabs) for 32 total, letting an agent provision hosting, databases, auth, and observability then bill it all through Stripe. New guardrails assign agent identities, enforce scope rules, and route sensitive actions through approval flows. The Projects skill installs into Claude Code, Cursor, or any MCP coding agent. This is procurement collapsing into a CLI, and the guardrails are exactly the kind of spend-control layer today's bankruptcy story screamed for.

Incumbents embed native agents to defend the seat. The same defensive playbook hit three categories in one quarter. Figma shipped an in-canvas design agent (revenue +46%) weeks after Anthropic's "Claude Design" knocked its stock ~7%. Tableau, Looker, and Power BI shipped autonomous BI agents that replace dashboards with "ask a question." Salesforce shipped Agentforce with credit pricing. Rather than get unbundled by standalone AI apps, the incumbents are racing to make the agent native to the system of record. That's the 2026 survival strategy, and it's why I'd bet against the simplest "AI eats SaaS" version of the story.

AI-native GRC tools move on Vanta, Wiz, and Snyk. Per Scytale, Delve auto-generates audit evidence, Scytale and Strac run continuous control monitoring, and FlowAssure automates vendor assessments across SOC 2, HIPAA, ISO 27001, GDPR, and PCI. The pitch is architectural: agents handle evidence collection end-to-end instead of the dashboard-and-checklist workflow Vanta pioneered. Compliance is a paperwork problem, and paperwork is exactly what agents are good at eating.

---

Policy & Governance

Anthropic reverses a covert Fable 5 safeguard: "We made the wrong tradeoff." After backlash over a hidden mechanism buried in Fable 5's 319-page system card, Anthropic reversed course June 11. The covert system silently degraded Claude for frontier-LLM-development queries using prompt modification, steering vectors, and parameter-efficient fine-tuning. Flagged requests will now visibly fall back to Opus 4.8 instead. The lesson for builders is concrete: invisible model interventions are a reputational third rail, and you should read system cards for silent capability gates. If a model quietly gets worse at a specific task and you don't know why, the answer might be in a 300-page PDF nobody reads.

EU rejects Apple's Siri AI exemption. Per MacRumors, the European Commission publicly rebuffed Apple on June 9, saying its decision to withhold Gemini-powered Siri from 450M+ EU users is Apple's own business choice, not a DMA requirement. Apple had sought an 18-month exemption from DMA Article 6 interoperability rules. Regulators refused. The standoff over who has to open up assistant access keeps escalating.

OpenAI, Anthropic, and Google executives to attend the G7 summit. Per a French presidency list reported by Bloomberg, Sam Altman, Dario Amodei, and Demis Hassabis are slated to attend. The frontier-lab leaders are now physically in the room where global AI governance gets debated. Whatever you think of that, it's where the rules will get shaped.

---

Skills of the Day

1. Set a hard spend ceiling at the billing layer before any unattended agent run. Prompts don't stop loops, budgets do. Cap absolute spend in your provider dashboard and add a per-run token budget plus an iteration limit in your harness. Today's DN42 bankruptcy was a missing ceiling, nothing more exotic.

2. Replace RL prompt tuning with GEPA reflective evolution. Use dspy.GEPA to optimize prompts by having an LM reflect in natural language on execution traces. It beats GRPO by up to 20% with up to 35x fewer rollouts because it consumes text feedback, not just a scalar reward. Stop hand-tuning.

3. Split high-value generation into separate generate, critique, and synthesize calls. Route the work through distinct roles with clean contexts: one drafts, one critiques against explicit criteria, one synthesizes. Costs more, but on high-stakes outputs the quality lift is real because each role gets one objective instead of competing instructions. Willison's planning-with-Fable-5, building-with-GPT-5.5 split is the same idea across models.

4. Diagnose context failures with write/select/compress/isolate. Long-running agents fail four ways: poisoning, distraction, confusion, clash. Each maps to a lever. Write (offload to files), select (retrieve only what's relevant), compress (compact at a token threshold), isolate (hand to a subagent that returns a 1-2k summary). Bigger windows don't fix these. The right lever does. Arize.

5. Keep SKILL.md under 500 lines and link out. Treat the skill like a manual: a lean table of contents up top, detailed procedures in linked files Claude reads only when needed. Eight well-structured skills load ~500 tokens at startup versus ~70,000 if everything loaded eagerly. Progressive disclosure is a ~140x token win.

6. Configure a fallbackModel chain so a rejected turn retries instead of dying. Claude Code v2.1.166 takes up to three models tried in order and auto-retries once on the fallback when the API throws an unexpected non-retryable error. For scheduled agents this turns a transient hiccup into a silent recovery. Put cheaper models lower in the chain to bound cost.

7. Default RAG chunks to 512 tokens with 10-20% overlap, then tune by query shape. A 50-document 2026 benchmark found 512-token chunks with 50-100 token overlap scored highest as a default. Factoid lookups do well at 256-512, analytical and multi-hop at 512-1,024. Add hybrid retrieval plus reranking for a 25-40% precision lift over naive vector search. Chunk sizing is the cheapest knob with the biggest payoff.

8. Pin extended-thinking to zero on routine turns with MAX_THINKING_TOKENS=0. Claude Code v2.1.166 made disabling reasoning reliable. Switch it off for mechanical edits, formatting, and scripted runs where thinking only adds latency and spend. Reserve the budget for genuinely hard planning and debugging. For unattended scheduled agents, a direct lever on per-run cost.

9. Return interactive HTML from MCP tools with MCP Apps. The first official MCP extension lets a tool return a UI resource via the ui:// scheme; the host renders bundled HTML/JS in a sandboxed iframe with callServerTool and requestDisplayMode. Forms, dashboards, and multi-step approvals run in-conversation instead of as text. Already live in ChatGPT, Claude web/desktop, VS Code, and Goose. MCP blog.

10. Probe hidden states instead of generating text when you only need a signal. If all you want is an intent label, a safety flag, or a route, read the model's activations with a probe instead of paying for a full generation. Cheaper, faster, smaller hallucination surface. Stop asking for an essay to get one bit. j11y.io.

---