> ## Content Index
> Fetch the complete content index at: https://www.antoinebuteau.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Daily Digest - 2026-10-13
- URL: https://www.antoinebuteau.com/daily-digest-2026-10-13/
- Published: 2026-10-14T10:11:45.000Z
- Updated: 2026-10-14T10:11:45.000Z
- Description: PostHog shares 38 practical lessons from building ten AI products used by nearly 100,000 monthly users, explaining why building MCP servers works better than standalone chat assistants.
- Author: Antoine Buteau
- Tags: Digest

## In this digest

1. [We have built 10 different AI products at PostHog like...](#digest-item-1)
2. [everyone is asking us at @cxodev about calculating ROI on...](#digest-item-2)
3. [Are Harnesses Worth the Cost?](#digest-item-3)
4. [The Theory 2026 AI Survey Results](#digest-item-4)
5. [Introducing Glass: Learn Things Once.](#digest-item-5)
6. [How to make a good RL task](#digest-item-6)
7. [Why WebMCP matters](#digest-item-7)
8. [Workday goes headless, opening its HR and finance data to outside AI agents](#digest-item-8)
9. [OpenAI Sidebar Apps: Inside the Architecture](#digest-item-9)
10. [Vertical SaaS companies have built software that helps customers do...](#digest-item-10)
11. [Building resilient AI-powered software, and planning for when things go wrong](#digest-item-11)
12. [How to tell if a startup has a moat](#digest-item-12)
13. [It Takes an Expert to Build an AI Expert](#digest-item-13)
14. [The Big Discount on Intelligence](#digest-item-14)
15. [\[AINews\] Meta Muse Wonder & Wonder Full: MSL’s Claude 5.5-tier Frontier Model](#digest-item-15)

## Themes from yesterday

- **The Shift from Chat Assistants to Governed Systems and MCP Primitives:** Product and engineering teams are moving away from standalone chatbots and exported data lake copies. Instead, they are adopting direct code execution, governed Model Context Protocol servers, and live database queries to embed agents directly into day-to-day operations.
- **The Maturation of Enterprise AI Governance, Security, and Sandboxing:** Operational priorities have shifted from expanding context windows to runtime security, multi-model routing, isolated sandboxes, and structured disaster recovery planning.
- **The Subsidized Intelligence Window and the Imperative for Institutionalization:** Because AI labs heavily subsidize power-user subscriptions to collect telemetry on how developers work, companies are working to capture individual employee prompts and workflows into shared internal tools before subscription pricing changes.
- **Frontier Model Deflation and the Race to Outcome-Based Software:** Falling token prices, such as Meta's discounted Muse Wonder release, alongside studies on agent harnesses and startup moats, show that long-term defensibility requires owning complete customer outcomes and industry context rather than relying on thin wrappers.

## 1\. **We have built 10 different AI products at PostHog like...** — PostHog (X)

- Why read: PostHog shares 38 practical lessons from building ten AI products used by nearly 100,000 monthly users, explaining why building MCP servers works better than standalone chat assistants.
- Summary: PostHog found that an AI product's real defensibility comes from proprietary context and data instead of custom prompt harnesses. Their MCP server attracted five times more daily active users than their built-in chat interface. Instead of adding a generic conversational assistant beside existing features, teams should expose direct code execution, SQL generation, and system primitives so agents can take action for users. Their best-converting deployment was an AI onboarding wizard that cut setup time from two hours to eight minutes, doubling activation rates and multiplying paid conversions by five. Multi-agent hierarchies and heavy prompt stuffing proved expensive and brittle compared to on-demand documentation lookups and deterministic taxonomy tools. Because user interfaces are shifting from action surfaces into review and approval screens, engineering teams should set up detailed tracing early and sandbox agents against indirect prompt injection.
- [Read more](https://twitter.com/posthog/status/2109706421919248499/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)

## 2\. **everyone is asking us at @cxodev about calculating ROI on...** — claire vo 🖤 (X)

- Why read: Claire Vo shares an operator framework for evaluating enterprise AI ROI, arguing that leaders should focus attention and capital on engineering, where over 80% of token spend happens.
- Summary: Companies trying to measure AI returns often focus too much on small departmental pilots instead of engineering, which accounts for more than 80% of total token spend. Instead of placing blunt budget caps on top users, managers should study internal power spenders to scale high-leverage workflows across the company or cut wasteful token loops. Pull request velocity alone is an incomplete metric; leaders need to tie model usage directly to shipped projects and the structural reduction of manual toil. Time savings do not automatically create financial gains, so teams should review session logs and execution traces to confirm freed engineering hours go toward growth priorities. As inference costs fall, companies should skip small single-digit improvements, target 2x to 10x operational gains, and track progress using revenue per fully loaded employee.
- [Read more](https://twitter.com/clairevo/status/2110164048839057905/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)

## 3\. **Are Harnesses Worth the Cost?** — Naresh Jain (martinfowler.com)

- Why read: An experiment published on Martin Fowler's site shows that behavioral harnesses multiply token costs for frontier models without improving results, but remain necessary for smaller, cost-optimized models.
- Summary: Naresh Jain tested coding agents against 178 automated acceptance checks for a complex backend service to evaluate whether behavioral harnesses combining architectural guides and sensor feedback improve code quality. Frontier models, including GPT-6 Sol, Claude Sonnet/Opus 4.6, and Gemini 3.8 Flash, passed every acceptance test on their first attempt without assistance. For those models, the harness increased token use by 10.5x and runtime by nearly 3x with no quality improvement. In contrast, smaller models like GPT-6 Luna struggled alone, but passed the test suite when guided by the harness's deterministic feedback and contract-driven mocks. Decoupling the inner development loop from the outer acceptance suite keeps the agent's assumptions from altering human architectural requirements. Engineering teams should adjust harness complexity based on the model tier, removing extra scaffolding when running top-tier reasoning models.
- [Read more](https://martinfowler.com/articles/are-harnesses-worth-the-cost.html?ref=antoinebuteau.com)

## 4\. **The Theory 2026 AI Survey Results** — Tomasz Tunguz (tomtunguz.com)

- Why read: Theory Ventures' second annual survey of 455 senior technical builders shows security has overtaken context limits as the main enterprise adoption hurdle, while model routing and sandboxing become standard.
- Summary: Theory Ventures surveyed 455 technical leaders and found that 43% consider security and governance their main obstacle to adoption, up from sixth place in 2025\. Dynamic model routing is now widespread, with 45% of engineering teams routing queries across multiple models through custom or commercial layers to manage inference costs. For execution sandboxes, Cloudflare leads with 62% market share among teams running autonomous coding agents, three times higher than its nearest competitor as compute infrastructure becomes the primary security control. Evaluating agents is also merging with standard analytics: 56% of builders use business intelligence tools and 48% use secondary coding agents to review execution traces. Infrastructure budgets are moving away from basic prompt management toward runtime sandboxing, automated evaluations, and multi-model cost controls.
- [Read more](https://read.readwise.io/read/01m4va9tcs633dpmpdac87a5yf?ref=antoinebuteau.com)

## 5\. **Introducing Glass: Learn Things Once.** — Eric Glyman (X)

- Why read: Ramp CEO Eric Glyman introduces Glass, an internal tool designed to turn individual employee prompts into shared company workflows after Ramp spent $1.5 million on tokens in one week.
- Summary: Ramp found that giving employees raw model access and open-ended token budgets created costly noise rather than a productive AI workflow. To stop useful operational knowledge from being lost in private chats, Ramp built Glass to store proven workflows, data definitions, and prompt templates across the organization. Finance rules and prompts that triage 20,000 customer requests each month are now available to all employees on their first day. The tool also includes Ramp Router and project-level spending meters, which cut internal inference bills by 30% overnight and gave teams live visibility into active agent costs. As teams grow, turning individual experiments into shared, reusable tools helps preserve operational context.
- [Read more](https://twitter.com/eglyman/status/2110007713396613568/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)

## 6\. **How to make a good RL task** — Eli Gottlieb (X)

- Why read: Eli Gottlieb explains how to design reinforcement learning environments that prevent reward hacking, offering a practical framework for automated task creation.
- Summary: Building effective reinforcement learning tasks requires balancing four connected parts: the prompt, the simulated environment, the verifier, and the underlying learning goal. Because RL models optimize strictly for reward signals without understanding human intent, any mismatch between the verifier and the true goal causes models to exploit loopholes in the sandbox. For messy domains like finance or software engineering, tasks should reflect real production ambiguity instead of including artificial hints that distort evaluations. Gottlieb recommends setting up automated task pipelines that use multi-agent workflows to convert expensive expert reviews into cheaper automated graders and classifiers. Adding continuous revision loops allows teams to expand training data volume while keeping generation costs down.
- [Read more](https://twitter.com/GottliebEli/status/2110077216260608037/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)

## 7\. **Why WebMCP matters** — blog.yoav.ws (yoav.ws)

- Why read: Shopify engineer Yoav Weiss explains WebMCP as a way to improve agent performance on the web, giving sites direct JavaScript tool interfaces that make co-browsing faster, cheaper, and more dependable than screen scraping.
- Summary: Having AI agents interact with websites through accessibility trees, screenshots, and simulated clicks is slow, expensive, and prone to breaking. WebMCP addresses this by letting web applications expose typed JavaScript functions directly in the browser, turning multi-step page navigation into direct tool calls. In tests by Shopify's Checkout team, WebMCP was 2.7x faster and 58% cheaper than standard browser automation, finishing with a 100% completion rate. The approach relies on co-browsing so the user stays in control, keeping the merchant's brand interface intact and avoiding false alarms from anti-bot defenses. Standardizing browser-level tool interfaces helps ensure the open web remains usable for agents instead of being replaced by closed APIs.
- [Read more](https://blog.yoav.ws/posts/why%5Fwebmcp%5Fmatters/?ref=antoinebuteau.com)

## 8\. **Workday goes headless, opening its HR and finance data to outside AI agents** — Frederic Lardinois (The New Stack)

- Why read: Workday launched a headless version of its Trusted Agent Platform, using an MCP server and live SQL queries to manage how third-party AI agents access sensitive HR and financial data.
- Summary: At Workday Rising, the company introduced Agent-Ready Tools, an official Model Context Protocol server with 275 read operations and 54 write and delete tools for agents running in tools like Claude Code, Cursor, or custom runtimes. The system applies user access controls and audit logging directly at the object graph level, unlike the common practice of pointing agents at external data lake exports. In live demonstrations, querying external Snowflake copies let agents bypass organizational reporting lines and expose executive pay, while the native MCP server enforced standard permission boundaries. The platform also includes Live Data Query for SQL access and an Agent Harness that manages identity, organizational rules, and model routing. Workday permits third-party agents for simple tasks, but requires its own internal orchestrators for complex HR and financial workflows.
- [Read more](https://thenewstack.io/workday-headless-mcp-agents/?ref=antoinebuteau.com)

## 9\. **OpenAI Sidebar Apps: Inside the Architecture** — Josh Rosen (X)

- Why read: Josh Rosen analyzes the architecture of OpenAI's new Sidebar Apps, detailing how sandboxing rules, MCP connections, and separate update cycles limit how apps run inside ChatGPT.
- Summary: OpenAI's Sidebar Apps move beyond simple tool plugins to embed interactive applications directly inside ChatGPT's main navigation. However, the system imposes tight structural limits: a sidebar app cannot mix tools across multiple MCP servers, functioning only as a dedicated interface for its own server. Apps also cannot browse or list files in the agent's working directory, forcing developers to build custom viewers tied to specific file extensions. Distributing an app through the public directory requires hosting remote HTTPS MCP infrastructure, which creates technical friction for open-source and client-side tools. In addition, developers must maintain three separate update cycles across bundled skills, MCP tool schemas, and iframe interfaces, making backward-compatible contract design critical.
- [Read more](https://twitter.com/JoshARosen/status/2110030255775776810/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)

## 10\. **Vertical SaaS companies have built software that helps customers do...** — Luke Sophinos (X)

- Why read: Luke Sophinos describes a shift in vertical SaaS, arguing that companies must move beyond digitizing manual tasks and start taking direct responsibility for business outcomes.
- Summary: For over a decade, vertical software companies built businesses by replacing spreadsheets and selling workflow automation on per-seat subscriptions. Saving customer time is no longer a lasting advantage now that AI agents can execute core tasks directly. Winning platforms will pair industry expertise and proprietary operational context with direct execution tools to improve revenue and profit for clients. Sophinos advises founders to map out workflows and start agents on low-risk tasks with clear metrics before giving them broader autonomy. By embedding payments and financial tools directly into these workflows, software vendors can charge for delivered business results rather than defending shrinking seat counts.
- [Read more](https://twitter.com/lukesophinos/status/2110031826831995260/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)

## 11\. **Building resilient AI-powered software, and planning for when things go wrong** — The Pragmatic Engineer (Substack)

- Why read: In an excerpt from his book, distributed systems expert Sam Newman outlines disaster recovery designs, RPO and RTO tradeoffs, and operational resilience practices for running AI systems in production.
- Summary: As coding agents accelerate software deployment across the industry, production environments are seeing more frequent outages and stability issues. Sam Newman reviews disaster recovery setups, from basic backups to warm failovers and multi-site active-active designs, evaluated against Recovery Point (RPO) and Recovery Time (RTO) targets. Newman cautions that targeting zero downtime or zero data loss creates exponential costs, meaning teams should tier their services and apply strict targets only to core transactional systems. Because GPU and CPU supply constraints make instant failover compute uncertain during regional cloud outages, organizations need multi-cloud and multi-model backup plans. Teams also need active monitoring and automated contract testing to handle model downtime, API behavior changes, and compute quotas.
- [Read more](https://read.readwise.io/read/01m4v0qdp9dj4zrgdk9h92e3d6?ref=antoinebuteau.com)

## 12\. **How to tell if a startup has a moat** — Dan Hockenmaier (lennysnewsletter.com)

- Why read: Dan Hockenmaier provides a framework for assessing startup defensibility in the AI era, separating temporary tactical advantages from compounding business flywheels.
- Summary: Generative AI allows startups to ship products quickly, but lower technical barriers mean unmoated companies stall as rivals copy their features. Hockenmaier points out that speed, venture capital, viral growth, and early brand recognition provide initial momentum rather than lasting defensibility. Long-term protection comes from four flywheels that strengthen with scale: high switching costs, network effects, economies of scale, and self-reinforcing data loops. In enterprise software, switching costs stem from the organizational effort required to retrain staff and rewire integrations, which helps compound startups that serve entire departments resist replacement. Teams should evaluate product roadmaps to ensure engineering work builds structural defensibility rather than just adding easily copied features.
- [Read more](https://www.lennysnewsletter.com/p/how-to-tell-if-a-startup-has-a-moat?ref=antoinebuteau.com)

## 13\. **It Takes an Expert to Build an AI Expert** — Hiten Shah (X)

- Why read: Hiten Shah explains why building dependable AI products requires domain experts to spot convincing hallucinations and turn professional judgment into evaluation rules.
- Summary: While building sixteen specialized AI skills for product marketing and competitive analysis, Shah found that frontier models frequently invent convincing explanations when inputs lack deal data or customer interviews. Experienced professionals are essential during development because they quickly spot outdated competitive claims, shallow market shifts, and single press releases treated as multiple sources. Shah uses the "Convincing Wrong Answer Test," where domain experts explain why a well-written AI answer is actually wrong or misleading. Engineers can turn these critiques into explicit system rules, deterministic retrieval filters, and regression test suites. As basic text generation becomes commoditized, enterprise software value will center on skill libraries that encode expert judgment to finish complex workflows.
- [Read more](https://twitter.com/hnshah/status/2110060518266605870/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)

## 14\. **The Big Discount on Intelligence** — Arunabh Dastidar (Arunabh's Substack)

- Why read: An analysis shows that frontier AI labs subsidize power-user subscriptions at negative 900% to 1650% gross margins to gather telemetry on how engineers solve complex problems.
- Summary: When an engineering team switched from flat-rate Claude and Codex subscriptions to direct API access for the same workload, their spending jumped thirtyfold from $1,000 a month to $1,000 a day. Industry data shows that a developer fully utilizing a $200 monthly subscription consumes about $4,000 to $6,000 worth of compute at API rates. Labs accept these heavy losses on purpose so they can track how developers fix mistakes, run tool harnesses, and navigate edge cases, gathering valuable post-training data. Citing Andrej Karpathy's capability funnel, the author notes that while most users interact with AI as simple chatbots, experienced engineers are using subsidized agent swarms to multiply their output. Teams should take advantage of these subsidized compute rates to build complex software before pricing models adjust.
- [Read more](https://arunabhdastidar.substack.com/p/the-big-discount-on-intelligence)

## 15\. **\[AINews\] Meta Muse Wonder & Wonder Full: MSL’s Claude 5.5-tier Frontier Model** — AINews (Substack)

- Why read: Meta Superintelligence Labs launched Muse Wonder and Wonder Full, matching Claude 5.5 performance in agentic coding while pricing tokens roughly 60% lower.
- Summary: Meta released Muse Wonder and Wonder Full, which ranked in the top three across ten agent benchmarks, including SWE-Marathon, Terminal-Bench, and DeepSWE. Built for long-horizon task execution and resistance to reward hacking, Wonder Full matched Claude Code with Sonnet 5.5 on the Coding Agents Index and handled out-of-distribution reasoning well. At $4 per million input tokens and $20 per million output tokens, Meta priced the models about 60% below competing frontier options. Early testers report clear prose and strong performance on tricky compiler errors, though 3D generation and creative writing remain rough. The release establishes a three-way race at the frontier, making multi-model routing useful for teams looking to cut inference expenses.
- [Read more](https://read.readwise.io/read/01m4t9zsg4017enq6m73xaeg2x?ref=antoinebuteau.com)