Agents’ Codex

Practical, no-hype writing on AI agents — cost control, multi-agent architecture, and real-world operations.

Independent publication, not affiliated with OpenAI.

Looking for OpenAI Codex how-tos? We’re building a dedicated series on working with Codex agents — check back soon.

MicroVM security isolation: a sealed glass chamber protects a fragile filament from a fractured iron plate, symbolizing agent isolation comparison

Coding Agent Sandboxes: Docker vs Brig MicroVM

Three isolation architectures now dominate coding agent security, and the real choice hinges on one question: whose kernel do you trust?

Glowing glass spheres linked by thin lines across a dark surface, suggesting a network of cooperating agents

OpenAI Codex Subagents: Run Parallel Multi-Agent Workflows

A practical guide to spawning and coordinating parallel Codex subagents, from TOML custom agents to the Responses API multi-agent beta.

Declarative agent configuration models compared: TOML, Markdown, YAML, A2A protocol

OpenAI Codex Custom Agents: TOML vs Markdown vs YAML

OpenAI Codex declares custom agents in TOML, SKILL.md and AGENTS.md use Markdown, Conductor uses YAML, and A2A skips manifests entirely. The right choice depends on whether you need enforceability, readability, or interoperability.

Cumora agent relay chain with mid-handoff gap illuminated, representing coordinated team execution

Cumora: Agents as First-Class Teammates in Team Chat

Cumora is team chat rebuilt so agents hold the same roster, DMs, and calendars as humans, not a chatbot bolted on top.

State-aware A2A routing decision node with three branching agent paths: self, collaborate, handoff

State-Aware A2A Routing: How Sprix SAGE Decides Mid-Task

The real gap in multi-agent systems isn’t discovery; it’s deciding who keeps working after execution has already begun.

Dedicated compute for AI coworkers: a matte-black Docker container as an industrial crate with glowing orange rectangular interface, isolated on dark concrete

OpenBot: Production Patterns for AI Coworkers with Dedicated Compute

OpenBot treats agent compute as a governance boundary: one container, one browser, and a fail-closed policy gate in front of every action.

Cost Regression Detection for AI Coding Agents

AI coding agents burn tokens invisibly; cost regression gates turn that waste into a failing build signal you can act on.

Agent identity governance as a vault of brass keys with one singled out under hard light - runtime credential control

Agent identity: runtime governance for non-human principals

Stop handing agents long-lived secrets; govern credential use the moment each agent acts, not at rest.

Diffusion LLM inference as hundreds of spheres resolving from fog all at once - parallel generation, not sequential

Diffusion LLMs for Agent Inference: Speed and Tradeoffs

A production engineer’s guide to when parallel token generation beats autoregressive decoding for latency-sensitive agent workloads.

Federated MCP as three separate stone platforms linked by light with no central hub - distributed tool access without a server

Federated MCP: Distributed Tool Access Without a Central Server

The hard part of federated MCP isn’t plumbing between servers; it’s deciding who can invoke what across a boundary no single party controls.