I founded Seed Club, a new model for early-stage investing built around networks, shared intelligence, and coordinated support. I’m interested in what happens when AI makes context, memory, and coordination more legible, and what that means for how companies and organizations get built.
My agent drafts this site from what I save. Green is me.
Week of May 11, 2026
Agents compose, persist memory, and self-evaluate through traces
The dominant architecture pairs a general orchestrator with specialists called as tool calls, while agents that write to memory after each task build compounding advantages. Reasoning traces, not raw compute, are the primary bottleneck for performant agentic systems.
- Forcing agents to render UI exposes errors they otherwise skip
- Traces enable automatic eval generation and self-debugging by other agents
- Running 100 continuous cloud agents signals post-scarcity developer tooling
Specialized experts like Harvey become tools the general agent pays to call · Nathan Baschez (@nbaschez) Agents that write to their own memory after every task compound an advantage · @gregisenberg OpenClaw runs 100 cloud agents reviewing every PR as if tokens are free · @steipete Reasoning traces, not compute, are the real bottleneck in agentic systems · Saneel (@sanlsrni) Agent traces let other agents auto-generate evals and debug themselves · ben hylak (@benhylak) Forcing agents to render a UI surfaces mistakes they would otherwise miss · @goodalexander
Agent-native tooling layer is forming and monetizing fast
Infrastructure designed specifically for agent consumption is taking shape: pay-per-use API catalogs, version control for agentic workflows, and multi-agent CLIs. Agents are becoming first-class economic actors with tooling needs distinct from human-facing software.
- Premium subscriptions can substitute for costly API calls in agent pipelines
- Version control for agents reached $35M market cap in two months
Hermes agents now search X via Premium instead of burning paid API calls · Kevin Simback (@KSimback) Git for agents Gitlawb hits $35M market cap in two months · @igoryuzo Grok Build runs a clickable CLI juggling many agents at once · Jason Ginsberg (@JasonBud) Ampersend launches a pay-per-use API marketplace built for AI agents · ampersend (@ampersend_ai) Claude debuts agent view, a unified dashboard for all coding sessions · Claude (@claudeai)
Venture fund sizing math is fracturing under round inflation
Series A inflation toward $30-100M forces dedicated seed funds to $300M+ or retreat to pre-seed, while megafund exit requirements have warped 'venture-backable' into a near-meaningless label. A $5B exit generates excellent returns for most funds but fails the pools that must clear $50B.
- Secondaries culture pushes founders toward tradable value over durable companies
- Anthropic voiding named secondary transfers signals active corporate counter-pressure
- Sub-$5M new-category bets outperform large raises in crowded proven markets
If Series A hits $30-100M, seed funds need $300M+ to stay relevant · @nunzi46 Selling-to-secondaries advice pushes founders toward tradable over durable · @arian_ghashghai Sub-$5M new-category bets beat $100M raises in crowded markets · Neil (@neilhar) Altos organizes funds around companies, heretical in venture · Ho Nam (@honam) Anthropic declares all secondary transfers of its stock illegal and void · Gabriel Shapiro (@lex_node) Venture-backable is meaningless when only megafunds need $50B exits · Arian Ghashghai (@arian_ghashghai)
AI dissolves execution moats, elevates taste and trust
AI has eliminated traditional product development bottlenecks, collapsing the execution gap that once separated high-performers. What remains scarce is taste, strategic judgment, and the trust built through genuine human presence in networks.
- VC analyst work is replaceable by agents; curated human networks are not
- Whether data or integration moats hold depends entirely on team execution
- The outlier founder rarely works on the consensus hot theme of the year
The hot theme rarely produces the year's most valuable company · Hadley Harris (@Hadley) Permanence is the rarest asset class in an age of AI disruption · Saneel (@sanlsrni) AI eats VC analyst work, leaving trust and network to humans · Rebecca Kaden (@rebeccakaden) AI startup moats range from real to false depending entirely on the team · Deedy (@deedydas) AI deleted the execution moat and pushed the constraint up to taste and judgment · Scott Belsky (@scottbelsky)