/ About this event
What Instinct and Muse ask of the agent stack—and where Jev’s System One approach could fit.
​About the Event ​What does it take for a personal agent to keep working toward your goals throughout the day?
​Products such as Instinct and Meta’s Muse raise expectations for agents that understand personal context, remember preferences, work across apps, and follow through as circumstances change.
​Delivering that experience requires continuous decisions: Which update matters? Is the information current? Should the agent act, think more deeply, or ask a person? Those choices compound across tasks and days of ongoing assistance, shaping responsiveness, reliability, user trust, and cost.
​TypeSafe AI’s Jev, a “System One” model designed for fast, structured decisions, brings an architectural possibility into focus: how could a specialized decision layer work alongside deeper reasoning, retrieval, memory, tools, and runtime to support personal agents at that frequency?
​Join SEAMATE × CORESPEED in downtown San Francisco for a founder and builder mixer connecting personal agent experiences with the architectures behind them. Through a keynote, panel, and builder showcases, we’ll explore what these choices could enable—and where product and infrastructure opportunities may emerge.
​Come to compare architectures, share lessons from what you’re building, and meet potential collaborators.
​One Personal Agent, Five Connected Layers ​Consider a personal agent coordinating a trip around a changing work calendar: noticing updates, retrieving live options, remembering constraints, proposing revised plans, and seeking approval before committing changes. This workflow connects five layers of the stack:
​Search — What does the agent need to know before deciding? Could fast decisions identify when to check the calendar or retrieve fresh travel options? How do retrieval and source verification supply evidence for both quick decisions and deeper planning?
​Decisions — Which choices need Jev-style speed, and which need deeper reasoning? Classifying an update or choosing an allowed next step differs from replanning an itinerary. Where could Jev fit, and how should confidence and validation guide escalation to a reasoning model or a person?
​Memory — What makes a fast decision personal and relevant? Budget, preferences, commitments, and corrections affect the next step. How should agents update stale context and give both decision and reasoning models the right information?
​Tools — How does a decision become an authorized action? MCP, APIs, or browser interactions carry out the proposed booking change. Which permissions must be enforced in code? How do valid structured outputs, correct decisions, and successful actions differ?
​Runtime — How does the agent keep working as circumstances change? How should persistent state, sandboxes, retries, and observability support frequent, lightweight decisions and deeper planning across background work and interrupted tasks?
​Across all five layers, we’ll evaluate task success, decision accuracy, confidence calibration, human interruptions, latency, and cost per successful task. When do faster decisions improve the experience, and when does retrieval or execution remain the bottleneck?
​What This Could Change ​Connecting the expectations raised by Instinct and Muse with approaches such as Jev’s System One decisions opens three product questions:
​More useful initiative: If frequent decisions become affordable and fast, what could an agent proactively notice or prepare—and when is an interruption worth the user’s attention?
​Deeper delegation: How do decision quality, memory, permissions, and recovery together earn the trust needed for people to hand over longer-running goals?
​New places to build: If agents mediate access to apps, how might distribution change? Which advantages come from specialized context and workflows, and which capabilities could become shared infrastructure?
​Tentative Agenda ​All times are local to San Francisco (Pacific Time).
​5:00–5:15 PM — Check-in & Welcome
​5:15–5:30 PM — Opening Keynote: From Personal Agents to Decision Architecture
​How the demands of ongoing personal assistance lead to new choices across decisions, reasoning, search, memory, tools, and runtime.
​5:30–6:15 PM — Panel + Audience Q&A: Personal Agents, Fast Decisions & Real-World Tradeoffs
​Using Instinct and Muse to frame product expectations and Jev to examine a possible decision layer, we’ll discuss how architecture shapes initiative, trust, cost, and the opportunities for builders.
​6:15–7:00 PM — Agent Open Mic · 5-minute Builder Showcases
​Show a real agent workflow, explain how its decisions are made, and share where it succeeds or needs human intervention.
​7:00–8:00 PM — Founder & Builder Networking
​Speaker, partner-session, and demo details will be announced. The agenda may be updated as the lineup is finalized.
​Who Should Attend ​Founders building personal agents, specialized agent products, developer tools, and AI infrastructure
​Engineers working on retrieval, memory, orchestration, tools, and runtime systems
​Product and technical leaders designing experiences around delegation and ongoing assistance
​Investors exploring opportunities across personal AI and the agent stack
This event is a part of #SFTechWeek—a week of events hosted by VCs and startups to bring together the tech ecosystem. Learn more at www.tech-week.com
