Huan Li

United States

Community Contributions

LatestBlog postsOpen sourceSpeaking Articles

Belief vs Control: /proc and /sys as an Interface for Agents

🔍 Belief vs Control: What Linux’ /proc and /sys teach us about AI agent interfaces. In traditional operating systems, /proc gives you a view into what the kernel believes, while /sys is the authority surface where control lives. In our latest post, we argue that AI-native agent runtimes need the very same split — not only to make introspection human-friendly but to separate observable state from mutation control. PromptWareOS adopts this design with /proc as belief surfaces and /sys as control knobs for agents — bringing introspection, safety, and evolution to agent systems without accidental mutation or brittle contracts. 👉 Dive deeper: https://ship.fail/blog/2025/12/20/belief-vs-control-proc-and-sys-as-an-interface-for-agents/ If you’re building AI co-founders, multi-agent systems, or operator-grade agent platforms, designing interfaces that clearly separate what the agent thinks from what you’re allowed to change is foundational to debuggability, trust, and long-lived behavior. 💡 #AIagents #AINative #AIarchitecture #PromptWareOS #DeveloperExperience

Blogpost / 01-21-2026

From Message Chaos to Order: How I Rebuilt the Mailbox Library for XState@5

If you’ve built event-driven systems, you’ve likely felt the pain of lost messages — when events arrive while an actor is busy, they simply vanish. In our latest post, I walk through how we rebuilt the Mailbox library for XState v5 to give actors true queued mailboxes — ensuring messages are processed reliably and in order. With native actor support in XState v5, we went from ~7,000 LOC hacks fighting the library to a clean, ~500-line mailbox implementation that wraps any machine with: ✔️ message queuing ✔️ sequential delivery ✔️ reply routing ✔️ observable subscriptions …and no more silent drops. This pattern matters for cloud-native, CQRS, and distributed systems: instead of losing commands or state transitions, your actor processes each event one at a time, predictably — like a queue in Erlang or Akka. 👉 Explore the full story and example usage: https://ship.fail/blog/2026/01/08/from-message-chaos-to-order-mailbox-xstate-v5/ #DevTools #EventDriven #XState #ActorModel #TypeScript

Blogpost / 01-09-2026

Why LLMs Can Copy UUIDs (and How That Makes Agents Safer)

Why can large language models copy long random strings like UUIDs character-for-character — even though generation is technically probabilistic? 🤔 In our latest post, we unpack this paradox and show how the same mechanism that lets LLMs perfectly repeat identifiers can be turned into a powerful safety anchor for AI agents. At a high level, models don’t “remember” tokens the way a database does — they use content-addressed retrieval and induction patterns in self-attention so that when a string appears in context, the next token distribution collapses around the correct continuation rather than spreading across all plausible tokens. This creates what looks like precise copying. We tie this in with Goodwin Checks — simple, automatable integrity probes that use exact string recall (e.g., invariants, IDs, constants) to detect when an agent loses anchors or drifts cognitively. When retrieval becomes unreliable (e.g., deep in context), these checks act like cognitive checksums that ensure the agent still “knows who it is.” 👉 Read the full post here: https://ship.fail/blog/2025/12/29/why-llms-can-copy-uuids/ #AIagents #AINative #ModelIntegrity #PromptWareOS #AgentSafety

Blogpost / 12-30-2025

We Built Wooden Headphones: How We Stopped Cosplaying Linux and Started Building an OS

When we started PromptWare ØS, we copied the form of Linux — “kernel,” “Ring 0,” “syscalls” — hoping that alone would bring OS-grade stability. But like islanders carving wooden headphones to summon planes, we were performing rituals without real mechanisms. That’s cargo-cult thinking — adopting structure without understanding the underlying constraints that make it work. In our latest post, I share how we faced that uncomfortable mirror, drafted RFC 0002: No Cargo Culting, and rewrote our architecture so names reflect real behavior and invariants. • Renamed misleading concepts (e.g., exec() → shell() because it actually runs shell commands). • Kept only mechanisms that enforce constraints (syscall boundaries, VFS abstractions, PID ordering). • Removed metaphor layers that gave a false sense of “OS-ness.” 👉 Read how we stopped cosplaying Linux and started building an actual AI-native OS: https://ship.fail/blog/2025/12/27/wooden-headphones-cargo-culting/ Building real systems means avoiding surface imitation and grounding design in actual mechanisms, not names. 💡 #AIarchitecture #PromptWareOS #CargoCult #SystemDesign

Blogpost / 12-28-2025

We Tried to Make Our Prompts Shorter. We Made Them Longer Instead. Somehow, That Made Them Better.

In traditional systems, “shorter is cheaper.” But *in an LLM-native OS where language is code, efficiency is about information density, not brevity. 🚀 In our latest essay, we introduce the STOP Protocol (Semantic Token Optimization Protocol) — born from a simple realization: 👉 abbreviations like op, ctx, and ref cost the same tokens as full words like operation, context, and reference in modern LLM tokenizers. That insight flipped our optimization strategy on its head. STOP flips the optimization script: ✔️ If two forms cost the same tokens, choose the one with more semantic clarity. ✔️ Standard conventions stay — ID, URL, API — because they mean something. ✔️ Token cost becomes a resource discipline like CPU cycles or memory in traditional kernels. This isn’t “brevity for savings.” It’s semantic maximalism — maximizing human meaning per token so PromptWare OS v0.5 boots faster, runs cleaner, and stays comprehensible. Whether you’re building kernel-level prompts or agent systems, the STOP Protocol shows that in LLM stacks clarity is efficiency. 👉 Read the origin story and principles here: https://ship.fail/blog/2025/12/26/stop-protocol-origin-story/ #AINative #PromptWareOS #TokenEngineering #SemanticClarity #LLMDev

Blogpost / 12-27-2025

PromptWare v0.5: Enforcing Cognitive Integrity with Software Physics

🧠 Prompts are suggestions. Physics enforces reality. In the v0.5 update of PromptWareOS, we take a major step beyond “vibes-based” AI engineering by introducing software physics — a system of immutable constraints that prevents autonomous agents from forgetting who they are or violating their identity contracts. Prompts alone are soft instructions — useful, but easy for a model to drift away from as it hallucinates or loses track of context. To build a true operating environment for AI co-founders, PromptWare v0.5 uses cryptographically isolated identities backed by persistent store (Deno KV), so an agent literally cannot act unless it knows who it is. Key v0.5 advances: ✔️ Goodwin Check: Agent identity enforced at the runtime layer, not in prompts. ✔️ Supervisor Layer: Every system call must pass identity verification before execution. ✔️ Microservices Architecture: Clear separation of concerns with atomic, stateless tools. ✔️ Persistent kernel parameters (/proc/cmdline) replace brittle memory in context windows. Read the full update and see how deterministic engineering locks in cognitive integrity for long-lived agents: https://ship.fail/blog/2025/12/21/promptware-v0-5-enforcing-cognitive-integrity-with-software-physics/ #AINative #AIarchitecture #PromptWareOS #SoftwarePhysics #AgentOrchestration

Blogpost / 12-22-2025

Ship.Fail Crew: The First PromptWareOS Distribution

Operating systems are more than kernels — they’re userlands, shells, utilities, and conventions that make machines usable. The same applies in the AI era. 🚀 In our latest post, we introduce Ship.Fail Crew — the first distribution for PromptWareOS, designed to be mounted as the AI co-founder’s home directory. If PromptWareOS is the kernel for prompt-driven systems, then Crew is the userland your agents actually run in. Crew gives you: ✔️ A mountable $HOME of agents, skills, and tool contracts ✔️ Shared conventions so promptware boots consistently across repos ✔️ A clean separation between kernel, distribution, and your project worktree Instead of copy/pasting sprawling prompt text from repo to repo, you mount one distribution and let every project boot into the same curated environment. Think of it like Debian for AI co-founders — a stable foundation your promptware can depend on. 👉 Read the full launch post: https://ship.fail/blog/2025/12/19/ship-fail-crew-the-first-promptwareos-distribution/ #AIarchitecture #PromptWareOS #AINative #AIagents #SoftwareDistribution

Blogpost / 12-20-2025

Booting Intelligence: The Architecture of PromptWar̊e ØS v0.2

In AI-native systems, intelligence shouldn’t disappear as context windows fill up. In our new article, we reveal PromptWar̊e ØS v0.2 — an evolved architecture that treats the OS as the keeper of memory, identity, and stability, rather than leaving it to the ephemeral context of an LLM. 🚀 In v0.2, we tackled two core problems: ✔️ Context Drift — when an agent forgets its “root” and hallucinations creep in. ✔️ Memory Fragility — the context window is not persistent. Key architectural advances: • A Kernel Memory Subsystem backed by Deno KV that locks in identity and state outside the sliding context window. • A strict separation between immutable bootloader spec and mutable state. • A tool-based context model: local tools operate on files; OS calls use a clear virtual filesystem (os://…). • A Just-In-Time linker that hydrates minimal YAML/Markdown source into rich runtime context only when needed. 👉 This work marks a shift from simulation-style agents to systems with physics: stable, composable, and reliable. https://ship.fail/blog/2025/12/18/booting-intelligence-architecture-promptware-os-v0-2/ If you care about long-lived AI co-founders and predictable multi-agent systems, this is the OS that makes intelligence boot — and stay booted. 💡 #AINative #AIarchitecture #PromptwareOS #AgentOrchestration #SoftwareDesign

Blogpost / 12-19-2025

Related Stars

Angie Jones

angiejones

Gleb Bahmutov

bahmutov

Chip Huyen

chiphuyen

Colby Fayock

colbyfayock