惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

钛媒体:引领未来商业与生活新知
钛媒体:引领未来商业与生活新知
U
Unit 42
GbyAI
GbyAI
M
MIT News - Artificial intelligence
美团技术团队
罗磊的独立博客
雷峰网
雷峰网
量子位
博客园 - 【当耐特】
Last Week in AI
Last Week in AI
D
Docker
小众软件
小众软件
S
SegmentFault 最新的问题
Blog — PlanetScale
Blog — PlanetScale
阮一峰的网络日志
阮一峰的网络日志
宝玉的分享
宝玉的分享
T
Tailwind CSS Blog
WordPress大学
WordPress大学
V
V2EX
博客园_首页
腾讯CDC
The Cloudflare Blog
A
About on SuperTechFans
Cyber Security Advisories - MS-ISAC
Cyber Security Advisories - MS-ISAC

DEV Community

Authentication Security Deep Dive: From Brute Force to Salted Hashing (With Java Examples) Why AI Systems Don’t Fail — They Drift Spilling beans for how i learn for exam😁"Reinforcement Learning Cheat Sheet" I Replaced Chrome with Safari for AI Browser Automation. Here's What Broke (and What Finally Worked) How Python Borrows Other People's Work The $40 Architecture: Processing 1 Billion API Requests with 99.99% Uptime Vibe Coding: A Workflow Guide (From Zero to SaaS) Most webhook security guides protect the wrong side. The scary part is delivery. Headless CMS for TanStack Start: Build a Blog with Cosmic EU Age Verification App "Hacked in 2 Minutes" — What Actually Happened Comfy Cloud’s delete function does not actually remove files Running AI Models on GPU Cloud Servers: A Beginner Guide Event-driven media intelligence with AWS Step Functions and Bedrock I scored 500 AI prompts across 8 quality dimensions — here's what broke How to Call Google Gemini API from Next.js (Free Tier, No Backend Needed) The Portal Protocol: Reclaiming Human Connection in the Age of AI How to Fix Your Team's Scattered Knowledge Problem With a Self-Hosted Forum Intro to tc Cloud Functors: A Graph-First Mental Model for the Modern Cloud Designing Multi-Tenant Backends With Both Ownership and Team Access I Built a Neumorphic CSS Library with 77+ Components — Here's What I Learned PostgreSQL Performance Optimization: Why Connection Pooling Is Critical at Scale Cómo construí un SaaS multi-rubro para gestionar expensas en Argentina con FastAPI + Vue 3 🚀 I Built an Ethical Hacking Scanner Tool – Open Source Project I Replaced /usage and /context in Claude Code With a Single Statusline A Pythonic Way to Handle Emails (IMAP/SMTP) with Auto-Discovery and AI-Ready Design I Collected 8.9 Million Polymarket Price Points — Here's What I Found About How Markets Really Move EcoTrack AI — Carbon Footprint Tracker & Dashboard Everyone's Using AI. No One Agrees How. 5 self-hosted ebook managers worth trying in 2026 Building Your First AI Agent with LangChain: From Chatbot to Autonomous Assistant
BoxAgnts (5) — Rust Is Becoming the Infrastructure Langua...
Guyoung Studio · 2026-05-29 · via DEV Community

Over the past two years, the AI Agent ecosystem has changed dramatically.

In 2023, most projects were still fundamentally:

LLM API + Prompt + Tool Call

Typical examples included:

  • LangChain
  • AutoGPT
  • CrewAI
  • AutoGen

These projects mainly focused on:

  • Prompt orchestration
  • Tool routing
  • Multi-turn conversations
  • Workflow orchestration

In essence, they were still:

LLM Workflow Frameworks

But starting in 2025, a major shift began to emerge:

Agents started gaining real permissions.

Agents were no longer limited to:

  • Calling APIs
  • Returning text responses

They began to:

  • Execute shell commands
  • Modify codebases
  • Access local file systems
  • Run continuously
  • Schedule multiple tools
  • Maintain persistent memory
  • Operate autonomously on local machines

This fundamentally changed the center of gravity in AI Agent engineering.

The core question shifted from:

How should prompts be designed?

to:

How should the runtime operate?

And this is precisely where Rust started entering the AI Agent infrastructure layer.


The AI Agent Ecosystem Is Clearly Layering

By 2026, the AI Agent ecosystem is no longer “Python everywhere.”

The field is increasingly stratified:

Layer Representative Projects Dominant Language Primary Goal
Prompt / Workflow Framework LangChain, CrewAI, AutoGen Python LLM workflow orchestration
IDE / Coding Agent Cline, Continue, Claude Code TypeScript IDE-integrated development workflows
Agent Runtime / CLI Agent OpenAI Codex CLI, BoxAgnts Rust Local runtime + tool execution
Personal Agent Platform OpenClaw TypeScript + native runtime Long-running personal agents
AI Inference Backend Tabby Rust Local inference and code completion

This distinction matters.

Because:

Rust is not competing with LangChain.

Rust is entering a completely different layer:

Runtime / Sandbox / Capability Security

That is an entirely different engineering domain.


Why Agent Runtimes Are Starting to Need “Systems Languages”

Many people still assume:

“Agents are basically just API wrappers.”

That may be true for simple chatbots.

But a real agent runtime must handle:

  • File system access
  • Shell execution
  • Multi-tool concurrency
  • Long-running tasks
  • WebSocket streaming
  • SSE parsing
  • Local indexing
  • Permission management
  • Resource limits
  • Sandboxing
  • Multi-agent scheduling
  • Tool capability management

These problems are much closer to:

Operating Systems / Runtime Engineering

than traditional web application development.

Especially once agents gain:

  • Autonomous execution
  • Local code execution
  • File system permissions
  • Persistent memory
  • Background execution

security rapidly becomes the core issue.

Recent research has increasingly focused on:

  • tool poisoning
  • memory poisoning
  • capability escalation
  • long-running agent attack surfaces

The community is beginning to realize:

The hardest problem in agents is no longer intelligence — it is permissions.


Why Rust Is Entering the Agent Runtime Layer

Rust’s real strength is not prompt orchestration.

Rust excels at:

  • Runtime systems
  • Sandboxing
  • Capability isolation
  • Resource metering
  • Deterministic execution
  • Multi-agent scheduling
  • WASM runtimes
  • Tool execution
  • Systems-level concurrency

In other words:

Rust is better suited for building the “operating system” of agents rather than the “glue layer” for prompts.

This is also the fundamental difference between BoxAgnts and traditional Python agent frameworks.

BoxAgnts is closer to a:

Capability-Oriented Agent Runtime

rather than a:

Prompt Workflow Framework

Its focus is on:

  • Secure tool execution
  • Agent isolation
  • Runtime scheduling
  • WASM resource control
  • Capability authorization
  • Tool sandboxing

These are fundamentally systems engineering problems.


OpenAI Codex CLI: A Landmark Moment for Rust Runtimes

One of the clearest signals that Rust is entering the agent runtime layer is:

OpenAI Codex CLI moving toward a Rust runtime architecture.

This matters enormously.

Because the primary challenges of Codex CLI are no longer:

  • Prompt design
  • Tool registration

Instead, they are:

  • Local runtime management
  • Tool execution
  • Terminal session handling
  • Approval systems
  • Long-running agent loops
  • Autonomous execution
  • Local permission control

These are clearly runtime problems:

Agent Runtime

rather than traditional agent framework concerns.

Rust’s strengths align naturally with this environment:

Rust Capability Benefit for Codex CLI
Single binary distribution Simplified CLI deployment
No runtime dependency Stable local installation
tokio concurrency Efficient tool scheduling
Static linking Cross-platform deployment
No GC pauses Smooth terminal interaction
WASI / sandboxing Command isolation
Low resource usage Long-running execution

This is not accidental.

It reflects a broader trend:

Agent runtimes are becoming increasingly Rust-based.


The Limitation of Python Agents: Great for Workflows, Not for Runtimes

Python remains one of the most powerful languages in AI.

But its strengths are primarily:

  • Rapid prototyping
  • ML ecosystems
  • Prompt experimentation
  • Data processing
  • Workflow orchestration

rather than:

  • Runtime systems
  • Isolation
  • Sandboxing
  • Resource control

The reasons are fairly straightforward.


1. Python Runtimes Are Heavy

A typical Python agent startup process looks like this:

Start interpreter
→ Load virtualenv
→ Import dozens of dependencies
→ Initialize runtime
→ Start services

Real startup times commonly reach:

3–10 seconds

This is acceptable for web services.

But for:

  • CLI agents
  • Desktop agents
  • Local runtimes

it noticeably impacts usability.

Rust startup is effectively:

./boxagnts

and the runtime is ready.

There is no:

  • VM
  • JIT
  • import storm
  • virtual environment

This significantly improves:

Out-of-the-box usability


2. Python’s Concurrency Model Is Not Ideal for Runtimes

Python asyncio works well for:

  • IO workloads
  • Network services

But its limitation is:

CPU-heavy work blocks the event loop

Agent runtimes frequently need to perform:

  • indexing
  • embeddings
  • parsing
  • AST analysis
  • tool execution

These are not purely IO-bound tasks.

Python typically requires:

  • multiprocessing
  • thread pools
  • celery
  • external worker systems

to compensate.

Rust tokio naturally supports:

Mixed IO + CPU scheduling

For example:

tokio::spawn(async move {
    // async IO
});

tokio::task::spawn_blocking(|| {
    // CPU-heavy work
});

This becomes increasingly important for:

Long-running agent runtimes


TypeScript: Excellent for IDE Agents, Less Ideal for Runtimes

TypeScript has also been extremely successful in the agent ecosystem.

Especially projects like:

  • Cline
  • Continue
  • Claude Code

which heavily benefit from:

VSCode / Node.js / Web ecosystems

Its strengths include:

  • Massive npm ecosystem
  • Natural IDE integration
  • Unified frontend/backend development
  • Web-native UX
  • Mature async IO

But TypeScript runtimes also have inherent limitations.


Node.js Is Fundamentally a Single-Threaded Event Loop

Node.js works extremely well for:

  • Network IO
  • Web services
  • IDE extensions

But problems emerge once agents begin:

  • Running continuously
  • Scheduling large numbers of tools
  • Executing local code
  • Running indexing systems
  • Managing memory systems

Particularly because:

CPU-heavy tasks block the event loop

This is one reason many next-generation agent runtimes are shifting toward:

  • Rust cores
  • Native runtimes
  • Hybrid architectures

tokio: The Real Foundation Behind Rust Agent Runtimes

Many developers assume:

Rust async ≈ Node.js async

But they are fundamentally different.

tokio is much closer to a:

High-performance runtime scheduler

It provides not only:

  • async IO

but also:

  • work-stealing schedulers
  • multi-threaded runtimes
  • task migration
  • lock-free channels
  • structured concurrency

This means an agent runtime can simultaneously handle:

- SSE token streams
- WebSocket pushes
- indexing
- background tasks
- tool execution
- memory updates

without blocking the primary runtime loop.

This is one of the largest architectural differences between Rust runtimes and Node.js runtimes.


WASM: The Ideal Sandbox for Agent Tools

One of the hardest problems in AI agents is:

How can tools execute safely?

Traditional solutions typically involve:

  • Docker
  • subprocesses
  • virtual machines

But these are often too heavy.

WASM is emerging as a compelling alternative because it naturally provides:

  • Memory isolation
  • Capability-based access
  • Deterministic execution
  • Resource metering
  • Fast startup

These properties align extremely well with AI agent tool runtimes.


Rust + WASM: Host and Sandbox Consistency

Rust has a uniquely important advantage here:

The host runtime and WASM tools can both use the same language.

This creates enormous engineering benefits.


Unified Type System

Host runtime:

pub trait Tool

WASM tool:

impl Tool for ReadFileTool

No need for:

  • DSLs
  • Duplicate RPC schemas
  • Multi-language bindings

Unified Async Model

Host:

tokio async

Tool:

tokio async

No need for:

  • JavaScript bridges
  • Python bridges

Unified Toolchain

cargo build --target wasm32-wasip1

directly compiles sandboxed tools.

This is a major advantage for Rust in the emerging world of agent sandboxing.


Capability Security: The Next Major Runtime Problem

Many people still underestimate this shift:

The core challenge of agent runtimes is no longer the model.

It is:

Capability Security

Because agents increasingly possess:

  • Shell access
  • File system access
  • Network permissions
  • Persistent memory
  • Autonomous execution abilities

This effectively makes agents resemble:

User-space autonomous programs

And runtimes must begin solving:

  • Permission boundaries
  • Tool isolation
  • Resource quotas
  • Memory access restrictions
  • Network restrictions

These are fundamentally operating-system-level concerns.


Rust’s Real Advantage: Not “Faster,” but “More Controllable”

Many discussions about Rust focus on:

  • Performance
  • Zero-cost abstractions
  • No garbage collector

All of those are true.

But for agent runtimes, the more important advantages are:

Capability Rust Advantage
Predictable resource usage Stable long-running execution
No GC pauses Reliable real-time interaction
Static linking Simplified deployment
Strong type system Clear runtime boundaries
Memory safety Reduced runtime crashes
Native WASM support Natural tool sandboxing
Deterministic builds Reproducible runtimes

Agent runtimes increasingly resemble:

User-space micro operating systems

And this is precisely the domain where Rust excels.


Conclusion: Rust Is Entering the AI Agent Infrastructure Layer

Python still dominates:

  • Prompt engineering
  • Workflow orchestration
  • AI research

TypeScript dominates:

  • IDE agents
  • Browser agents
  • Web-native agent UX

But Rust is increasingly entering a new layer:

  • Agent runtimes
  • Sandboxing
  • Capability security
  • WASM isolation
  • Deterministic execution
  • Resource metering

This is not language replacement.

It is a structural evolution:

AI agents are becoming layered like operating systems.

As agents begin to:

  • Run continuously
  • Operate with real permissions
  • Execute local code
  • Access file systems
  • Schedule multiple tools
  • Maintain persistent memory

the runtime itself becomes the central problem.

And that is precisely where Rust is strongest.

Related Resources