notessamuele onelia

Dan Shipper

Dan Shipper is the co-founder and CEO of Every, a media and software company building agent-native applications (Reader, Anecdote, and others) alongside an editorial operation focused on the AI frontier. He writes the Chain of Thought column and hosts the AI & I podcast. Together with Claude, he authored the Agent-Native Architectures playbook, which synthesizes the principles his team has learned from shipping agent-first apps.

Key Themes

  • Agent-native design. Building applications where the agent is a first-class citizen with parity to the UI, not a chatbot bolted on.
  • Files as the universal interface. Bash + filesystem is the most fluent agent surface; everything else is built on top of that.
  • Latent demand through the agent. Watching what users ask the agent to do is the cleanest product research available.

Agent-Native Architectures (Jan 2026)

Why Now

Why Agent-Native Software Works Now — the conditions that make this category buildable in 2026

Core Principles

Parity - Agents Need Tools for Everything the UI Can Do — foundational principle Granularity - Tools Are Atomic Primitives, Features Are Outcomes — tools stay atomic; features are prompts Composability - New Features Through New Prompts — ship features by writing prompts Emergent Capability - Agents Do What You Didn’t Design For — the architectural payoff Improvement Over Time Without Shipping Code — the third payoff: compounding without releases

Tool Design

From Primitives to Domain Tools — the staging path Domain Tools Should Be Shortcuts, Not Gates — the constraint that protects composability Graduating Hot Paths to Optimized Code While Preserving Parity — the optimization path

Files-First Design

Files Are the Universal Interface for Agents — why files dominate as the agent interface Entity-Scoped Directories for Agent-Native Apps — a directory shape that’s working The context.md Pattern for Agent Memory — portable agent memory in one file Files for Legibility, Databases for Structure — when to choose each Conflict Models When Agents and Users Share Files — the trade-offs you take on by going files-first

Execution Patterns

Explicit Completion Signals Beat Heuristic Detection — agents need to say “done” explicitly Model Tier Selection by Task Complexity — right-size the model to the task Partial Completion with Task-Level Tracking — per-task state survives interruption Design Tools for Bounded Context Windows — context is finite; design for it

Implementation Patterns

Shared Workspace Over Separate Agent Sandboxes — agent and user in the same data space Context Injection in the System Prompt — tell the agent what exists and what it can do No Silent Agent Actions - UI Reflects Activity Immediately — visibility builds trust

Product Implications

Progressive Disclosure - Simple to Start, Endlessly Powerful — Excel-like product shape Latent Demand Discovery Through Agent Usage — agents as a continuous product research instrument Approval Stakes vs Reversibility Matrix — a starting heuristic for unsolicited agent actions Self-Modification Must Be Legible — the safety frame for self-modifying agents

Mobile

Checkpoint and Resume for Long-Running Agents on Mobile — surviving iOS interruptions iCloud-First Storage with Local Fallback for Mobile Agents — free cross-device sync Background Execution Within iOS Time Limits — the ~30-second discipline

Advanced Patterns

Dynamic Capability Discovery Over Static Tool Mapping — the most-atomic version of tool design CRUD Completeness Audit for Every Entity — the cheapest parity insurance

Anti-Patterns and the Test

Agent-Native Anti-Patterns - Router, Workflow Tools, Defensive Design — the named failure modes The Ultimate Test - Outcomes You Didn’t Build a Feature For — the single test that distinguishes agent-native from agent-flavored

See Also

Backlinks