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 25, 2026
Deterministic orchestration outperforms emergent multi-agent coordination
Structured, explicit workflows driving small agent loops are proving more reliable than non-deterministic multi-agent systems for complex tasks. New platform features and DSLs with primitives like parallel() and pipeline() are making this architecture the practical standard.
- Explicit workflow primitives replace emergent coordination between autonomous agents
- Structured orchestration unlocks multi-step tasks previously too complex for agents
- Context engineering, not model capability, is the remaining differentiator
Deterministic workflows orchestrating small agent loops beat agent soup every time · dex (@dexhorthy) New DSL gives subagents explicit workflow code instead of emergent coordination · Michael Livs (@micLivs) The /workflow feature cracks multi-step tasks agents long failed at · Jason Zhou Deterministic workflows around small agent loops beat multi-agent soup · dex Claude Code auto-generates orchestration plans it follows across hundreds of agents · cat
Infrastructure, not model quality, is now the real constraint
Model capability has crossed a threshold where harness, connectors, and reliable uptime matter more than raw intelligence. With nonhuman traffic exceeding half of some APIs, the build question has shifted from 'can the model do this' to 'can the system stay up and integrate.'
- No unified control plane yet exists for multi-domain agent orchestration
- Products lacking headless interfaces risk repeating the early mobile-era mistake
- Agents are beginning to manage their own threads, worktrees, and orchestration
Agents now need only a harness, connectors and reliable uptime, not better models · @nickbaumann_ No single control plane exists to manage agents across code, design and ops · Dennison Bertram Codex now manages its own threads and spins up worktrees for parallel tasks · Guinness Chen Agents are the next mobile, with 53% of one API's traffic already nonhuman · Gokul Rajaram Shopify's River is an AI agent living inside company Slack · shopify.engineering Sepo agent runs in GitHub Actions to implement issues and review PRs · github.com
Professional networks are becoming AI-queryable deal infrastructure
A set of tools now treats relationship graphs as structured, searchable data for finding warm backchannels, sourcing founders, and activating sales. The workflow collapses what previously required hours of manual research into a single natural-language query.
- Warm intros are closing more stuck deals than cold outreach
- Early founder signals surface from follower graphs before pitches arrive
- Network search now reaches across teammates' and friends' connections
Agents complement humans rather than replace them near-term
Spending on agents shows sharply diminishing returns beyond a baseline, while the competitive split is between people using AI and those who are not. The optimism and judgment required to act under uncertainty remain human advantages.
- Returns to agent expenditure fall off faster than for human labor
- The competitive divide is humans with AI versus humans without
- Intense conviction about outcomes remains a human edge agents cannot replicate