Week of November 16, 2026
OpenAI now wants to run your agent harness for you, plus the case that Sequoia's evergreen fund was the right idea launched at exactly the wrong time for its LPs.
This Week's Top Videos
Harness-as-a-Service: Building with the Agents API | DevDay 2026
By OpenAI
OpenAI learned teams don't want to build agent harnesses, they want OpenAI to run them. The Agents API is a managed Codex harness: you define tools, data, instructions, and environment, while OpenAI handles remote sandbox execution, observability for unpredictable agent paths, and state that survives deploys over hours-long tasks. Computer use over the API also ships now.
- Developers repeatedly asked OpenAI for opinions on agent infrastructure (sandbox placement, secrets management, tooling) and ultimately wanted OpenAI to run the harness. Knowing your workflow does not mean you should operate the harness yourself.
- OpenAI frames its stack as layers: the Responses API (2025) is the base interface for agentic harnesses, Codex adds 'hands' for coding, presentations, and browsing, and the Agents API sits on top as a managed service.
- Model capability changed the infrastructure problem. GPT-4 Turbo ran tool workflows lasting seconds in 2023, while GPT-6 Astra uses dozens of tools for hours or days, which forced a more flexible foundation than the Assistants API.
- Moving a laptop-native harness to the cloud means decoupling the harness from the execution environment. Every former local command must be routed to the correct session's sandbox and returned fast while keeping sessions isolated. You can bring your own sandbox or use an OpenAI-hosted one.
- Raw tool calls and model responses were not enough to build a good agent UX. Because agent paths are unpredictable, users need visibility into what the agent is doing and whether it is stuck, especially on long-running tasks.
- Computer use is now available over the API in OpenAI-hosted environments, letting Astra operate GUIs, not just terminals. Durable state must also survive server failures and code deploys during multi-hour tasks.
Sequoia Leaves London, Close Scout Program & End Evergreen Fund | NASDAQ Loses $2.5T in 24 hours
By 20VC
Sequoia killing its evergreen fund matters more than its London exit or scout shutdown. The structure was intellectually right, since post-IPO holding is where the real money compounds, but it launched in 2022-23 before a roughly 30% drawdown, trapped LPs with different liquidity needs, and made partners who were already paid force LPs to 'take their vitamins.'
- Sequoia's evergreen fund was built on a sound thesis. Sequoia-backed companies have made up something like 25% of the NASDAQ, so holding them after IPO would have compounded well above the S&P. It failed on execution, not logic.
- Timing sank the evergreen structure. Sequoia implemented it in 2022-23, right before stocks fell about 30%, and forcing LPs to hold through that drawdown hurt even if a 20-year horizon would smooth it out.
- Check incentive alignment before adopting any fund structure. Partners had already earned their carry, so forcing LPs to keep holding stock added risk for LPs with no extra upside for the GP. Individual partners can still hold their own shares forever. One Sequoia partner on Nvidia's board since 1993 is likely worth billions that way.
- Continuation funds are a cleaner way to hold generational assets. Altos Ventures uses them so LPs can choose to exit or stay, instead of being locked into a single evergreen vehicle.
- Founders should treat VC firms as durable and individual partners as temporary. Sequoia lost its CFO-turned-leader, all of its London partners, and several US partners. The founder-CEO, as with Jensen Huang, usually outlasts everyone on the cap table.
- Price IPOs conservatively. The panel argues it is far better to start low and trade up than to start high and come back, which is directly relevant to how Anthropic and other AI giants approach going public.
Claude Fable 5.2: How I’d Make Money With It
By Greg Isenberg
Greg Isenberg argues Claude Fable 5.2 lets you turn hourly freelancing into fixed-price productized services, citing Anthropic's real freelance score jumping from 21% to 41%. He shares copyable prompts for 10-12 hour autonomous business audits, 100-niche sub-agent market research, rebuilding legacy spreadsheets as web apps, and P&L deal checks that flag suspicious numbers.
- Isenberg says Fable 5.2 can work 10-12 hours autonomously. He uses that to run a consultant-grade audit covering website, pricing, onboarding, emails, and the top 5 competitors, with a running log and a ranked list of the 10 changes most likely to make money, each backed by evidence.
- Sell outcomes, not hours. Isenberg cites Anthropic's blog reporting a real freelance score jump from 21% to 41%. In his demo, the model picked a $450 fixed-price equity analyst job on Upwork and did the work, flagging which parts a human should check and what to charge.
- Use parallel sub-agents for market research. His prompt spins up one sub-agent per niche across 100 niches. Each finds the top 3 complaints, what people pay today, and how hard the market is to reach, then rolls everything into one ranked table with the top 5 niches to build for.
- Rebuild 'ancient software' as a per-output service instead of billing $50-100 an hour. For example, take a 12-year-old spreadsheet a client runs their business on, reverse-engineer what it does, rebuild it as a simple web app, migrate the data, and write a one-page staff guide.
- Isenberg claims the model now says 'I don't know,' which makes finance and legal work more viable. His deal-checker prompt for buying a boring business marks every number in a seller's P&L as verified, unverified, or suspicious and writes the 15 questions to ask before making an offer.
- Build a quality loop around the model for repeatable freelance work. Isenberg predicts thousands of people will package these jobs as fixed-price, job-to-be-done offers rather than hourly gigs.
24 Hours with Special Offer Founder & Art Director, Brent David Freaney I Day in the Life
By Figma
Brent David Freaney rebuilt his design studio Special Offer around a product mindset instead of a service one, packaging album art, logos and identity as a signature 'Special Offer product.' He picks clients who want his brutal, typographic aesthetic, runs a no-music 30-minute walk to collect references, and art-directed Madonna's cover through physical acetate layering before rebuilding it in Figma.
- Reframe a creative service business as a product business. Special Offer stopped acting as a vendor that 'helps get you the thing' and started selling a defined Special Offer product made of album art, logo and visual identity.
- Filter clients for fit, not volume. The studio deliberately works only with clients who want its specific brutal, typographic, maximal-minimal aesthetic, which protects quality and brand consistency.
- Treat mutual trust as the core asset in collaborative work. Freaney's method is to stay authentic to his style while translating it into each client's language, so both sides take pride in the output.
- Use your commute as a deliberate mental warm-up. Freaney walks about 30 minutes to the studio with no music, because lyrics push him into work mode too early, and he uses the time to collect visual references from street signage.
- Mine undesigned vernacular sources for authenticity. Truck graphics, bureaucratic signage, Chinatown supply shops and a restaurant doily scanned into a vinyl label all beat polished 'designer-looking' work.
- Prototype physically, then digitize. For Madonna's cover he spent two weeks of evenings at her house layering printed acetate typography and color sets, photographed the results, and rebuilt them in Figma.
The Rise Of AI Coworkers
By Y Combinator
Grockbot went from first line of code to launch in seven weeks by organizing work around always-on AI teammates with their own identity, phone number, credit card and computer, not around chats or files. Its performance marketing bot now runs Google and Facebook ads with its own login and budget, signaling how non-engineering functions get automated next.
- Grockbot is built on three principles: work is organized around bots rather than chats or files, bots get the same tools a human uses (including their own computer, not just MCPs), and bots are always on and proactive instead of waiting to be prompted.
- Trust escalates in stages. Users start with a basic task, then hand the bot an outcome or objective, then give it a full job once it proves reliable.
- The internal performance marketing team gave its bot its own Google Ads and Facebook Ads logins plus a budget, and had it coordinate with the design team's bot to generate assets and A/B test them. This is bot-to-bot collaboration across functions.
- Internal adoption followed a bell curve. Both the most AI-forward employees and some of the least AI-forward employees loved it, which suggests the teammate metaphor lowers the onboarding barrier.
- For engineers, Grockbot handles the 'outer loop' or 'last mile.' It brings agents into Slack threads and X mentions, so context from bug reports and feature requests no longer has to be summarized and pasted into a coding agent by hand.
- You can structure bots to match how your brain works. Split them vertically by function (engineer bot, marketer bot) or horizontally by entity (one bot per customer).
Rethinking Claude Code Best Practices — Lydia Hallie, Anthropic
By ai.engineer
Auto mode blocked 89% of clearly dangerous commands while only 14% of human reviewers caught them, yet about 1 in 10 still slipped through. Anthropic's Claude Code team now treats developers as managers of dozens of long-running sessions, which means lean CLAUDE.md files, plain-English auto mode settings, and hard deny rules for anything irreversible.
- Anthropic went from about 2,500 code changes a day in May to more than 30,000 a day in September. Engineers now run dozens of parallel Claude Code sessions for hours, and some sessions orchestrate other sessions.
- Strip your CLAUDE.md down to only what Claude cannot see from the code, such as a legacy service that is never directly referenced. 'IMPORTANT!' and 'NEVER do this' style instructions now push newer models off track, causing extra tool calls and worse results. Anthropic also ships a prompt audit command that flags what a newer model no longer needs.
- Auto mode runs a separate classifier on anything beyond reading or editing project files. You can configure it in plain English in settings by listing trusted repos, buckets, and domains plus what to always allow or never do. That cuts approval interruptions for tasks like pushing to a deploy repo.
- In Anthropic's study, only about 14% of human reviewers caught a clearly dangerous command, while auto mode blocked 89%. The classifier beats humans at catching danger, but roughly 1 in 10 risky commands still got through.
- Deny rules are the only real guarantee because they are enforced in the harness before the auto mode classifier runs. A matching tool call never even reaches the classifier. Blocking reads of your .env file, for example, also stops workarounds like cat.
- Use a simple reversibility test to decide what becomes a hard rule: if Claude gets it wrong once, can you take it back? Leave bad code, which you can revert, to the model. Lock down anything irreversible, like discarding uncommitted changes or leaking secrets into transcripts.