Conviction Brief — May 2026

147 bookmarks · −2 vs prior month (149 → 147) · rolls up 4 weekly briefs

Prior month baseline (April 2026 conviction brief) was not generated. All conviction marks below open fresh positions — no MoM deltas are available. Where directional movement can be inferred from the corpus itself (theme acceleration/cooling within May's four weeks), it is noted as intra-month movement, not a MoM delta.

The Month in One Move

The corpus spent May building the operator layer — skills, harnesses, memory, and workflow orchestration as the new unit of work — while the highest-edge signal for this reader was a quiet, repeated pattern: AI-native services delivered as companies, not tools. The dominant volume was unambiguously AI-builder Twitter (Claude Code workflows, CLAUDE.md, agent harness engineering, Obsidian-as-second-brain), and it converged on a clear thesis: the prompt is dead, the procedure is the product. Karpathy's CLAUDE.md (82k stars by mid-month), Thariq's "Unreasonable Effectiveness of HTML" (34k bm — the single highest-engagement bookmark of the month), and Garry Tan's meta-prompting posts crystallized a shared language for how to make agents work. But the edge-relevant signal was thinner and ran in parallel: Mark Goldberg's "AI personal trainers for execs" ($1k/hr AI-native operators, 535 bm), Anthropic's "Claude for Small Business" launch (12.3k bm), Cofounder 2's "run an entire company with agents" (8.2k bm), Higgsfield's end-to-end agentic ad pipeline (8.7k bm), and Greg Isenberg's SF trip report — "3 billionaires are all buying SaaS companies and rebuilding them agent-first" (6.7k bm). These are the seeds of the AI-native services thesis that June would later crystallize. In May they are present, repeated, and correctly labeled as Opening — not yet a window. Conviction 65/100 for the month as a whole; the edge-relevant signal is forming but not yet actionable on its own.

Signal vs Noise

All four weeks fired. Marks are against the month's own signal flow.

Net: The month's loudest signal (harness engineering) is correctly parked as operator fluency with zero founding edge. The edge-relevant signal (AI-native services) is present but early. The corpus remains ~90% SF AI-builder discourse.

Conviction Marks

Opening all positions fresh — no April baseline to delta against. Intra-month directional movement noted where observable.

Memory & Context Architecture — open at 60/100 (intra-month: rising)

• Confirming: massive volume across all four weeks. Ashwin Gopinath's Company Brain series (Parts 3–4, interaction/action memory), a16z's System-of-Intelligence thesis, Ian Lapham's personal research engine, cyrilXBT's Obsidian vault (15.5k bm — highest bm in the entire corpus), Polymath Investor's reading-retention guide (20.2k bm, highest engagement on a memory post). The maturity gradient is already being articulated (indexing → traces → entity extraction → authority weighting → conflict resolution) before June formalizes it.
• Disconfirming (required): every example is personal/individual or US-enterprise. No SEA/India org-memory signal. The technical moat is the content (Obsidian vaults, Notion integrations), not a GTM wedge. The anti-thesis is live but not yet articulated in May — context windows could collapse "memory" back into "prompt."
• Net: forming, real capability gradient, but already converging on a technical moat. Edge-fit LOW. Not a Found target. Weights: volume/capability-gradient +50%, technical-moat-not-GTM −30%, no-regional-proof −20%.

AI-Native Services — open at 65/100 (intra-month: rising through the month)

• Confirming: Goldberg's AI-operator-for-execs, Cofounder 2's agent-orchestrated company (8.2k bm, 7.7k likes), Higgsfield's end-to-end ad pipeline (8.7k bm), Light Anchor's autonomous e-commerce brands (YC-backed), Anthropic's "Claude for Small Business" (12.3k bm — the explicit SMB services framing), and Isenberg's SF billionaire report (6.7k bm, 6.2k likes). The pattern repeats across weeks and is asserted by credible operators (YC, Goldberg at Chemistry, Isenberg after SF trip).
• Disconfirming (required): every named example is US/SF-framed. Zero bookmarks show this working in SEA/India SMB contexts. Services businesses carry human-delivery drag, trust/regulatory friction, and worse margins than SaaS. The arbitrage is asserted, not demonstrated, in the reader's actual market. No customer-side evidence anywhere.
• Net: highest-conviction edge-aligned thesis forming in the corpus, but it's a hypothesis in May, not a window. The labor-cost-to-software-margin arbitrage is the spine. Weights: operator/YC convergence +45%, frontier-internal proof +25%, edge-fit alignment +20%, no-regional-proof −10%.

Loops / Workflows / Orchestration — open at 55/100 (intra-month: stable)

• Confirming: dominant theme by volume. Garry Tan's meta-meta-prompting (11.6k bm), Addy Osmani's harness engineering (8k bm), rohit4verse's 300-agent guide (1.1k bm), simonlast's large-scale coding agent lessons (6.8k bm), daniel_mac8/sidbid's Dynamic Workflows in Claude Code (4.8k + 2.8k bm). The unit of work is the workflow, not the prompt — strongly reinforced.
• Disconfirming (required): no anti-thesis articulated yet in May (that arrives in June with "don't build agents for everything"). Verifier-as-bottleneck is implied but not yet named. The field is still in maximalist expansion. This is operator fluency, not a Found target.
• Net: dominant by volume, but operator-fluency thesis (use it to run the firm), not an investable edge. Edge-fit NONE. Weights: volume-still-dominant +60%, no-GTM-wedge −40%.

Consumer AI (SEA/India + wellness) — open at 35/100 (intra-month: thin, barely present)

• Confirming: tsurezure's SEA streaming chart showing local-pop rising over K-pop/Western pop in Indonesia/Philippines/Thailand/Malaysia/Singapore (3.4k bm, 11k likes — genuine cultural signal); GPT-Realtime 2.0 voice demos at end of month (Farza, 8.4k bm); Suekhim's Koji AI tutor for kids (10.8k bm, 12.3k likes — consumer/education); Gabriel's design engineering patterns (11.8k bm — prosumer/creator).
• Disconfirming (required): n≈3–4 and none are SEA/India consumer product signals — the streaming chart is cultural observation, the voice demos are US-framed, Koji is US consumer edtech. Wellness is essentially absent (one peptides post). The corpus barely sees consumer at all.
• Net: low-confidence open. The SEA cultural signal is interesting but not a product signal. Edge-fit HIGH if a real wedge appears; right now there's almost nothing to act on. Weights: SEA-cultural-observation +40%, voice-as-new-surface +30%, wellness-missing −30%.

Startup Wedges to Explore

Spread across archetypes deliberately. Two real wedges beat five padded.

1. AI-native managed services firm in one SEA/India vertical (AI-native services · domain operator + a few licensed humans · e.g. SME bookkeeping/tax compliance, or SME ad creative-as-a-service · Indonesia or India)

• Window FORMING: Goldberg's AI-operator-for-execs and Cofounder 2's agent-orchestrated company prove the pattern is being funded; Higgsfield's ad pipeline proves the craft is automatable; Anthropic's "Claude for Small Business" proves the platform wants SMB services use cases. But no one has shown this working in SEA/India SMB contexts yet — the arbitrage is asserted, not demonstrated.
• Edge-fit: HIGHEST (AI-native services, SEA/India). Found when a regional probe clears; Back a returning operator sooner.
• First probe to falsify: pick ONE repetitive SME workflow (ad creative, bookkeeping, lead-qual), run it through existing tools for 5 SEA/India SMBs, measure cost-per-completed-engagement vs. the human-only incumbent. Falsifier: if they won't pay a recurring monthly fee, or if regulatory/trust friction forces enough human review that margin collapses to consultancy levels, it's a services business not a software-margin one.

2. AI family/household operations assistant for Asian dual-income households (consumer/wellness · consumer-PM operator · WhatsApp-native · schedules, eldercare coordination, school/health logistics · India/Indonesia/Philippines)

• Window TOO EARLY but worth tracking: GPT-Realtime 2.0 (voice as new surface) and Koji (AI tutor for kids, 10.8k bm) are the closest consumer signals. The multi-generational household + domestic-help-coordination structure of SEA/India is a genuinely different wedge than the US nuclear-family framing. WhatsApp is the native distribution rail. But no one has built "Ollie for SEA" yet, and the corpus has zero consumer-side demand evidence.
• Edge-fit: HIGH (consumer SEA/India + wellness). Found if a sharp consumer founder; Back otherwise.
• First probe: WhatsApp-based concierge MVP for 30 dual-income Mumbai/Jakarta households, fully human-run behind the curtain. Falsifier: if retention past week 4 needs constant human ops with no path to automation, it's a concierge service, not a product.

The Next Category [MANDATORY]

Voice as an operating surface (GPT-Realtime 2.0). The last 72 hours of May produced a sharp spike: Farza's "control my computer with just my voice" (8.4k bm, 14.2k likes — highest like-count in the month), Greg Isenberg's "17 startup ideas that ONLY work because of what this model makes possible" (2k bm), and nicbstme's "Agent Stack For Automating My Personal Life" (3.6k bm). When a capability drop from a frontier lab (OpenAI's Realtime 2.0) immediately produces this much builder ideation in the same week, that's the corpus signaling a new surface forming.

• The capability + why conviction is high: real-time voice gives agents a zero-friction input modality — the front door to the memory/workflow layer without a keyboard or chat box. Conviction that voice becomes a persistent operating surface is high (the demos are real, the engagement is off-the-charts for the week, and it connects directly to the "personal OS" thesis running through the entire month).
• Why edge-fit is MED-HIGH today (blunt): this is closer to the reader's lane than world-models or foundation models. Voice-as-surface maps to consumer (SEA/India, WhatsApp-native voice), prosumer (creator tools), and AI-native services (voice-driven service delivery). The edge exists if the application is regional consumer/services, not if it's building the voice model itself (NONE).
• The edge you'd have to BUILD or BUY: you don't build the voice model (OpenAI/Anthropic own that). You build the application — the consumer-facing voice concierge, or the voice-driven service-delivery layer for SMBs. That window is Opening — the capability just dropped, the demos are this week, and no SEA/India application exists yet.
• Who is better positioned to found it: a consumer-PM operator who understands SEA/India voice habits (voice notes as primary communication in Indonesia/India) and can build on top of Realtime 2.0. For this reader: Found the application layer if a sharp consumer founder is available; Back an early team building it. The discomfort is lower than world-models — this is application-layer, not research-layer.

Relevance Radar

• "Company brain" / org-memory is the forming infra category — Gopinath's maturity model, a16z's System-of-Intelligence, cyrilXBT's 15.5k-bm Obsidian vault. Real capability gradient, but converging on a technical (not GTM) moat. Watch as an enabler for services plays; do not found into it. Edge-fit LOW.
• Agentic ad creative is already shipping as a pipeline — Higgsfield's end-to-end ad agent (8.7k bm) and Light Anchor's autonomous e-commerce brands (YC) prove the craft layer is automatable now. The durable wedge is the regional managed-service relationship, not the tool — the tool is open and replicable within a quarter.
• Chinese AI labs building efficiently with fewer resources — Nathan Lambert's lab-visit report (1.8k bm) is a macro flag: the cost floor for AI-native services drops faster when capable models come from resource-constrained ecosystems. Relevant as a tailwind for SEA/India services unit economics, not a founding target.
• Stablecoins / agentic commerce rails going live — Jonah's stablecoin business analysis, Base's "Agentic Economy," Stripe/Bridge/Tempo. Infra-level shift; relevant as a rail under consumer/commerce wedges, not a founding target itself.

Corpus Blind Spots

• The bookmarks are ~90% AI-builder Twitter (harnesses, skills, memory, workflows, tooling) and almost blind to the reader's own edges. Consumer is n≈3–4; SEA/India is nearly absent as primary subject despite being the reader's HIGHEST-edge geography. The reader is sampling SF AI-tooling discourse, not the markets they can actually win in.
• No demand-side or customer voice anywhere. Every services/consumer signal is a builder announcing a launch or a VC asserting a category — there is not a single bookmark from an SMB, a D2C brand, or an end customer saying what they'd pay for. The corpus measures supply-side hype, not pull.
• Wellness, the named edge archetype, is essentially missing — one peptides post (morellifit, 2.3k bm) is the closest hit. If wellness is a real lane, the follow-list isn't feeding it.
• The "AI-native services" thesis is entirely US-framed. Goldberg, Cofounder 2, Higgsfield, Light Anchor, Isenberg's billionaires — all SF. The SEA/India services arbitrage is the reader's highest edge, and the corpus has zero regional proof points for it.

Q1–Q5

• Will SEA/India SMBs pay recurring (not project) fees for AI-delivered creative/services? The entire Found thesis lives or dies on retention economics in price-sensitive markets — and the corpus has zero customer-side evidence. The pattern is asserted by SF operators; the regional proof is the reader's to generate.
• Where does the AI-native services margin actually land after regulatory/trust friction — software-margin (70%+) or dressed-up consultancy (20–30%)? Cofounder 2 and Higgsfield assert the former; nobody's shown the post-friction number in a regulated vertical.
• Is "company brain" a category or a feature? The maturity model (levels 1–5) is real, but a16z's System-of-Intelligence framing and cyrilXBT's Obsidian vault suggest it may collapse into the platform (Claude/Notion/Obsidian), not stand as a company. What's the defensible layer — conflict resolution? Authority weighting? Or is it just "stores your stuff well"?
• Does voice (Realtime 2.0) unlock a SEA/India consumer wedge that chat never could? Voice notes are the primary communication modality in Indonesia/India. If voice-as-surface is real, the consumer application for multi-generational households is a genuinely different product than a US chatbot. Is this the front door the consumer thesis has been waiting for?
• When does the SF-framed "AI-native services" pattern get regional proof? The thesis is forming in May and will crystallize in June (YC's explicit decacorn framing). The reader's edge is being there first in SEA/India — but the corpus gives no signal on timing. What's the tripwire: a local operator launching, a YC company expanding, or a frontier lab releasing a regional product?

Key Reads

• Marc Andreessen — Current AI custom prompt · 53,233 bm
• Thariq — Using Claude Code: The Unreasonable Effectiveness of HTML · 34,194 bm
• Mnimiy — Karpathy's 4 CLAUDE.md rules cut Claude mistakes from 41% to 11% · 24,168 bm
• Andrej Karpathy — structure your response as HTML · 21,549 bm
• Polymath Investor — how to remember everything you read · 20,228 bm
• Angelica Parente — Tufte skill for Claude charts · 16,651 bm
• CyrilXBT — How to Build an Obsidian Knowledge Vault · 15,470 bm
• Dep — Karpathy's CLAUDE.md hit #1 on GitHub with 82,000 stars · 13,728 bm
• Nico — Anthropic launches "Claude for Small Business" · 12,281 bm
• Gabriel — Design Engineering interaction patterns · 11,795 bm

Signal Brief — Week of 2026-05-03

48 bookmarks this week

This Week's Opening

The window forming is AI-native services delivered as companies, not tools — and theloudest version this week is autonomous e-commerce and agent-run ad creative. YC spotlighted Light Anchor building "fully autonomous e-commerce brands at the speed of compute" (55 bm). Higgsfield shipped an end-to-end agentic ad pipeline — Claude researches competitor ads, Higgsfield generates creatives, Meta MCP launches and scales winners, runs every Monday (8,716 bm). Cofounder 2 launched "run an entire company with agents" across engineering, sales, marketing, ops, and design (8,217 bm). And Mark Goldberg at Chemistry put a price tag on it: "AI personal trainers for execs — AI-native 22-year-olds charging $1,000+/hr to automate workflows. Bigger than the $50B gym trainer market" (535 bm). The pattern is the same: the craft is automatable, the delivery is being productized as a company, and nobody has shown it working in SEA/India yet. Archetype: AI-native services. Found the regional version; the craft layer is commoditizing this month. Edge-fit: HIGHEST.

Moving Now

NEW or MOVED this week only:

Wedge Sketches

• Done-for-you AI ad studio for SEA/India D2C brands — actionable now. Higgsfield proved the pipeline works; the craft is commoditizing this week. The defensible layer is the regional relationship + done-for-you delivery at a flat monthly fee under a local agency retainer. Cheapest probe: run 5 paid pilots for Jakarta/Mumbai D2C brands using existing tools only — no build. Falsifier: if they won't pay recurring monthly (vs one-off project pricing), the services thesis is wrong for this segment.
• AI-native operator-as-a-service for SEA/India SMBs — speculative. Goldberg's "AI personal trainer for execs" reframed as a regional managed-service: a firm that deploys AI-native operators (not tools) to run back-office functions for SMBs at a fraction of the human-only cost. Why now: Cofounder 2 and Light Anchor prove the orchestration layer exists; Goldberg proves the pricing model. Cheapest probe: take one repetitive SME workflow (bookkeeping, lead-qual), deliver it via AI-in-the-loop for 5 SMBs, measure cost-per-completed-task vs. human baseline.

What's Breaking

The Obsidian-vault-as-second-brain content wave (cyrilXBT 15.5k bm, browomo 11k bm, coreyganim 410 bm) is converging on a technical moat, not a GTM wedge — the exact opposite of what a Found/Fund reader needs. The capability is real (governed knowledge base as memory layer), but the content is about how to build your vault, not what company to build on top of it. Anything in the pipeline premised on "Obsidian vault as product" should re-anchor on the service or relationship layer above it.

Carry-Forward

Watch whether the AI-native-services pattern (Goldberg, Cofounder 2, Higgsfield, Light Anchor) produces an actual regional services company — someone selling agent-delivered creative, bookkeeping, or ops in SEA/India — versus staying as US-framed tooling and assertions. The transition from "SF operator asserts the thesis" to "Jakarta SMB pays recurring fee" is the highest-edge event to catch opening.

Key Reads

• Marc Andreessen — Current AI custom prompt · 53,233 bm
• CyrilXBT — How to Build an Obsidian Knowledge Vault · 15,470 bm
• Blaze — Second Brain in Obsidian, 3 trading ideas, $180K in 6 months · 10,996 bm
• GREG ISENBERG — Google's Design.md + AI Skills · 8,860 bm
• Higgsfield AI — How to run paid ads with an agent in 2026 · 8,716 bm

Signal Brief — Week of 2026-05-10

45 bookmarks this week

This Week's Opening

The window this week is "System of Intelligence" as the new company brain — and the quiet emergence of AI-native onboarding as a lived experience. a16z formally named the shift: "From System of Record to System of Intelligence — own the layer that pulls from the SoR, becomes the user's one-stop shop for context and action" (2,303 bm). Ashwin Gopinath extended his Company Brain series with the architecture framing: "Claude Managed Agents Point to the Next AI Infra Layer" (819 bm). And Jean-Michel Lemieux (Spellbook) posted the most visceral signal of the month: "Joined a new AI-native company this week. The laptop arrived, I logged in, and an agent basically took over from there — set up dev env, pulled repos, fixed dependencies, got permissions approved" (4,127 bm, 6,263 likes). That is not a tool announcement; it's a lived organizational experience of agent-native work. Meanwhile, Thariq's "Unreasonable Effectiveness of HTML" (34,194 bm — the month's highest-engagement bookmark) and Karpathy's HTML-output post (21,549 bm) crystallized the operator-level insight: LLM output should be structured artifacts, not chat text. Archetype: enabler/infra → AI-native services. Edge-fit: LOW for the infra itself; HIGHEST for the implication (agent-native companies are forming, and the services version is the regional wedge).

Moving Now

NEW or MOVED this week only:

Wedge Sketches

• Agent-native services firm for SEA/India SMB back-office — actionable. jmwind's onboarding post proves agent-native companies are real; a16z's System-of-Intelligence names the infra; Harvey-connector commoditization proves the wrapper layer is dying. The wedge: a firm that delivers back-office functions (bookkeeping, compliance, lead-qual) to SEA/India SMBs using agent-native workflows, at a flat monthly fee. Why now: every layer below the service (infra, models, wrappers) is commoditizing; the service relationship isn't. Cheapest probe: take one regulated workflow, deliver it agent-assisted to 5 SMBs, measure margin after human-review friction.
• Agent-facing commerce/checkout for SEA/India D2C — speculative. Greg Isenberg's "new buyer is an AI agent" + stablecoin rails (Jonah, Base) + Shopify UCP. If agents are buying, the SEA/India D2C brand that builds agent-native checkout (MCP, stablecoin-settled) captures a new channel. Why now: the rails are going live this month. Cheapest probe: add an agent-readable product feed + MCP checkout to one Jakarta D2C brand, measure agent-driven purchase volume.

What's Breaking

The Harvey-connector integration cuts against any wedge premised on vertical-AI-as-wrapper as a durable moat. If Anthropic ships 12 plugins doing what Harvey ($11B) does, for free, in the same release, then "legal-specialized AI on top of frontier models" is not a company — it's a feature of the platform. Anything in the pipeline that leans on "we have a vertical model wrapper" should re-anchor on service delivery, regulated trust, or encoded workflow IP instead. The craft layer commoditizes monthly; the relationship layer doesn't.

Carry-Forward

Watch whether a16z's "System of Intelligence" produces an actual company (someone owning the context+action layer for a vertical) versus collapsing into Claude/Notion/Obsidian as a feature. The transition from "category name" to "company with revenue" is the signal that the infra layer is investable — and it connects directly to whether the services wedge above has a durable memory/context moat or is just "runs Claude for you."

Key Reads

• Thariq — Using Claude Code: The Unreasonable Effectiveness of HTML · 34,194 bm
• Mnimiy — Karpathy's 4 CLAUDE.md rules cut Claude mistakes from 41% to 11% · 24,168 bm
• Andrej Karpathy — structure your response as HTML · 21,549 bm
• Nico — Anthropic launches "Claude for Small Business" · 12,281 bm
• Garry Tan — Meta-Meta-Prompting: The Secret to Making AI Agents Work · 11,636 bm

Signal Brief — Week of 2026-05-17

12 bookmarks this week

This Week's Opening

The signal this week is memory-as-moat crystallizing as a thesis — and the cost-of-compute question going mainstream. Two posts define the week. Polymath Investor's "how to remember everything you read" (20,228 bm, 13,574 likes — the highest-engagement bookmark of the entire month by likes) pushed the personal-memory protocol into virality. Ashwin Gopinath sharpened the architecture: "Rent the Intelligence. Own the Context." (565 bm) — the cleanest one-line thesis for why memory (context ownership) is the durable layer when model intelligence is rented. And Peter Steinberger's "people freaking out over my AI spend — what nobody sees: how would we build software if tokens don't matter?" (5,443 bm) opened the cost-engineering question that June would later formalize as "token engineering." Meanwhile, Greg Isenberg identified the compounding advantage: "Hermes agents write to their own memory after every task. Starting today versus starting in 6 months is an unfair advantage" (2,928 bm). Archetype: enabler → AI-native services. Edge-fit: LOW for the memory layer itself; HIGHEST for the implication (cost-engineered services with governed memory = the regional unit-economics engine).

Moving Now

NEW or MOVED this week only:

Wedge Sketches

• Governed-context services firm for SEA/India — speculative. Ashwin Gopinath's "own the context" + Isenberg's "agents write to their own memory" + Shreyas's "deep product context" converge: the firm that builds governed client/workflow context today has a moat that competitors can't replicate by just running the same models. For a SEA/India services firm, the context (client data, regulatory patterns, domain workflow history) IS the switching cost. Cheapest probe: start accumulating structured context for one vertical's workflow (e.g. Indonesia SME tax compliance), measure whether the context improves task-completion quality over 4 weeks vs. cold-start.
• Cost-engineered services as the bankable unit — speculative. Steinberger's "what if tokens don't matter?" is the enabling question. If token cost approaches zero, a SEA/India services firm can deliver at software-margin (70%+) what used to require human labor (20–30% margin). Cheapest probe: take one repetitive SME workflow, run it through a cost-engineered stack (routing, caching, cheaper defaults) for two weeks, measure cost-per-completed-task vs. the human baseline. If the margin clears 50% after friction, the thesis banks.

What's Breaking

The reading-retention content wave (Polymath Investor 20k bm) cuts against any wedge premised on personal-memory-as-product — the viral content proves demand exists, but it also proves the content is free. If "how to remember everything you read" is a 20k-bm tweet, the productized version has to be dramatically better than a free thread, or it's just a content business. Anything in the pipeline premised on "personal memory app" should re-anchor on either (a) governed org-memory for services (B2B, where the buyer pays for compliance/audit) or (b) a consumer concierge where the memory is a feature, not the product.

Carry-Forward

Watch whether Steinberger's "what if tokens don't matter?" produces an actual cost-engineering discipline (routing, caching, cheaper defaults as named practices) versus staying as a design philosophy. The transition from "philosophy" to "discipline with metrics" is the signal that the services unit-economics are bankable — and it's the precursor to June's "token engineering" thesis crystallizing.

Key Reads

• Polymath Investor — how to remember everything you read · 20,228 bm
• Dep — Karpathy's CLAUDE.md hit #1 on GitHub with 82,000 stars · 13,728 bm
• Dickie Bush — 8 hours reading 50+ AI articles, 8 you should read · 6,467 bm
• Peter Steinberger — "how would we build software if tokens don't matter?" · 5,443 bm
• Michael Morelli — Everything you need to know about peptides · 2,343 bm

Signal Brief — Week of 2026-05-24

19 bookmarks this week

This Week's Opening

The window this week is Dynamic Workflows shipping as a platform feature — and the "buy dead SaaS, rebuild agent-first" thesis getting billionaire confirmation. Anthropic shipped Dynamic Workflows in Claude Code (sidbid, 2,754 bm; daniel_mac8, 4,844 bm) — the platform now autonomously detects complex tasks, writes orchestration scripts, and spawns agent swarms. Simon Last (Notion) posted the highest-signal operator lessons: "running coding agents on large-scale projects — most of this contradicts advice from 6 months ago" (6,783 bm). And Greg Isenberg returned from SF with the thesis-confirming report: "I had lunch with 3 billionaires. All of them are buying SaaS companies and rebuilding them agent-first" (6,682 bm, 6,225 likes). Meanwhile, Angelica Parente's Tufte-skill post (16,651 bm — second-highest bm of the month) crystallized the skills-as-IP thesis: feed a book to Claude, get a reusable skill, instantly better output. Archetype: AI-native services + prosumer/creator. Edge-fit: HIGHEST for the SaaS-rebuild-as-services thesis; MED-HIGH for the skills-as-IP layer.

Moving Now

NEW or MOVED this week only:

Wedge Sketches

• Buy-and-rebuild: acquire/partner with SEA/India services firms, rebuild delivery agent-first — actionable now. Isenberg's billionaire report confirms the thesis; Dynamic Workflows shipping means the orchestration layer is free; token-budget wars mean the cost floor is dropping. The wedge: identify a local services firm (bookkeeping, legal docs, ad creative, compliance) with an existing customer base in Indonesia/India, acquire or partner, rebuild the delivery layer with agents. Why now: the billionaires are doing this in SF this month; the regional version hasn't started. Cheapest probe: identify 3 dead/stagnant local SaaS or services firms with recurring revenue, model the cost of rebuilding delivery with agents vs. their current human cost.
• Domain-skill-IP studio for SEA/India verticals — speculative. Parente's Tufte skill (16.7k bm) proved that encoded taste/judgment as a skill produces dramatically better output. The wedge: a firm that builds and owns domain-specific skills (e.g. Indonesia SME tax compliance skill, India lending ops skill, SEA insurance claims skill) and licenses the verified workflow as a service. Why now: the skill layer is API-accessible and the platform (Claude Code) supports it natively. Cheapest probe: hand-build one vertical skill, test whether generic-loop + your skill beats generic-loop alone by enough to charge for.

What's Breaking

Dynamic Workflows shipping as a platform feature kills any wedge premised on "we orchestrate agents" as the moat. If Claude Code autonomously detects complex tasks, writes orchestration scripts, and spawns swarms — natively, for free — then the orchestration layer is commoditized. The durable layer moves up: it's the encoded workflow knowledge (skills-as-IP), the domain context (governed memory), and the service relationship (regional trust). Anything in the pipeline that leaned on "we build agent orchestration" should re-anchor immediately.

Carry-Forward

Watch whether the "buy dead SaaS, rebuild agent-first" thesis produces an actual regional transaction — someone acquiring a SEA/India services/SaaS firm and rebuilding with agents — versus staying as a SF billionaire pattern. The transition from "SF billionaires do it" to "Jakarta/Mumbai operator does it" is the highest-edge event to catch opening. Also watch whether GPT-Realtime 2.0's voice surface produces a consumer application — if voice is the new front door, the SEA/India consumer concierge wedge (Ollie-for-SEA) may have just found its input modality.

Key Reads

• Angelica Parente — Tufte skill for Claude charts · 16,651 bm
• Simon Last — running coding agents on large-scale projects · 6,783 bm
• GREG ISENBERG — SF trip: billionaires buying SaaS, rebuilding agent-first · 6,682 bm
• GDP @ COMPUTEX — DeepSeek's 10 trillion USD grand strategy · 3,101 bm
• Garry Tan — YC as a Witness · 1,369 bm