AI Agents

AI agents are software that acts on your behalf—some wait for your direction, others run continuously in the background. The difference matters for security, cost, and control.

This section covers the full agent stack: autonomous platforms that operate independently, coding assistants that amplify your workflow, and the underlying language models that power them.

Three Ways to Think About Agents

Agentic Tools — Autonomous Platforms

Software that runs independently, makes decisions, and takes action without waiting for approval. Always-on systems that integrate into your workflows.

When you want this: Background automation, scheduled tasks, ambient intelligence

Key platforms:

• OpenClaw — Self-hosted, local-first agent platform with extensible skills
• Moltbook — Social network where autonomous agents post, vote, and coordinate

Read first: Architecture risk analysis — autonomous agents create new security boundaries

Coding Tools — Human-in-the-Loop Assistants

AI that works alongside you during development. You remain in control—the AI suggests, explains, and helps implement, but you review and approve every change.

When you want this: Writing code, debugging, refactoring, learning unfamiliar codebases

Key tools:

• OpenAI Codex — Cloud-native parallel agents with worktree isolation

Read first: Codex risk analysis — cloud agents create dependency and latency tradeoffs

LLM Models — The Foundation

The underlying language models that power both agentic tools and coding assistants. Understanding model capabilities, pricing, and benchmarks helps you choose the right engine for your agent.

When you want this: Selecting the right model, comparing benchmarks, understanding pricing

Current leaders:

• Claude Opus 4.5 — 80.9% SWE-bench, maximum reasoning capability
• Gemini 3 Flash — 78.0% SWE-bench, 1M context window
• Kimi k2.5 — 76.8% SWE-bench, best value with free access

Read first: Model comparison by tier — match models to your budget and performance needs

Agent Configuration

AGENTS.md — One File, Consistent Behavior

AGENTS.md (and its Claude Code equivalent, CLAUDE.md) is the single highest-leverage configuration for AI agents. One markdown file tells agents your conventions, commands, and guardrails—loaded automatically before every session.

Why it matters:

• Stops repetition — Document rules once, agents remember them
• Better than skills — Vercel’s evals show AGENTS.md beats skills on formatting reliability
• Universal — Works across OpenAI Codex, Claude Code, Cursor, and more

Start here:

• /agents/agents-md/ — Practical guide with templates
• /posts/agents-md-practical-guide/ — Evidence-based deep dive

Quick Decision Framework

Security Starts Here

All agent types create new attack surfaces. The risk model differs by category:

Browse all security analysis: /risks/

Recent Analysis

• Codex vs Claude Code vs Kimi k2.5 — Three paradigms for AI-assisted development compared
• OpenClaw Security Reality — What self-hosted actually means for your threat model
• Fetch-and-Follow Risk — When platforms can remotely control your agents

The Ecosystem at a Glance

┌─────────────────────────────────────────────────────────┐
│  APPLICATION LAYER                                       │
│  ┌─────────────┐  ┌─────────────┐  ┌─────────────────┐  │
│  │ OpenClaw    │  │ Moltbook    │  │ OpenAI Codex    │  │
│  │ (self-hosted)│  │ (social)     │  │ (cloud)          │  │
│  └──────┬──────┘  └──────┬──────┘  └────────┬────────┘  │
│         │                │                  │           │
│  ┌──────┴────────────────┴──────────────────┴────────┐  │
│  │           MODEL LAYER (LLM APIs)                   │  │
│  │  Claude Opus 4.5 • Gemini 3 Flash • Kimi k2.5     │  │
│  └────────────────────────────────────────────────────┘  │
└─────────────────────────────────────────────────────────┘

Stay Current

The agent landscape changes weekly. All content here includes last_reviewed dates and evidence levels. Check verification for fact-checking methodology.