Sign InOpen Brain
AI EngineerVideoSource Linked

Codex, Behind the Harness — Dominik Kundel, OpenAI

Codex exposes several reusable harness patterns: deferred tools, persistent browser execution, sandboxed edits, automated permission review, WebSockets, and server-side compaction.

AI Engineer · Aug 10, 2026
Open Source Open MarkdownOpen JSON
Source Summary

Codex’s open-source Rust harness caps skill descriptions at **2% of the context window** and defers some tools until search finds them. Since **GPT-5.4**, deferred tool loading is also available through the Responses API.

Practical Implication

For custom coding agents, borrow the boundaries rather than rebuilding every mechanism: persistent execution for browser scripts, apply-patch for edits, OS-specific sandboxes, read-only review agents for escalations, and server-side compaction for long runs.

Agent-Ready Context
Codex’s open-source Rust harness caps skill descriptions at **2% of the context window** and defers some tools until search finds them. Since **GPT-5.4**, deferred tool loading is also available through the Responses API.

For custom coding agents, borrow the boundaries rather than rebuilding every mechanism: persistent execution for browser scripts, apply-patch for edits, OS-specific sandboxes, read-only review agents for escalations, and server-side compaction for long runs.

At **1,000 tokens per second**, GPT-5.3 Codex Spark made network round trips the bottleneck, motivating a persistent WebSocket mode. These details describe the current harness and API surface; both are expected to shift with later models.
Connected Context · Feed7 Judgment

This turns several broad harness recommendations into concrete Codex boundaries: budget skill discovery, defer tools, preserve execution, constrain edits with sandboxes and review, and compact server-side. It reinforces progressive loading, isolation, and long-run context controls in the candidates, while adding a latency consequence: once inference is extremely fast, connection architecture becomes part of agent performance. The mechanisms remain model- and API-version dependent.

Skills are new features: Building Skill-Centric Harness — Yogendra Miraje, FactSetFactSet frames progressive skill loading and sandboxed execution as governance requirements; Codex provides concrete implementations through a context budget, deferred discovery, OS sandboxes, and review escalation.Evolution of agentic surfaces — Gagan Bhat & Isabella Kai He, AnthropicBoth show that harness behavior is coupled to model releases; Anthropic’s obsolete reset workaround is direct evidence for retesting Codex’s current compaction and loading choices as models change.Anthropic's CCA Exam as a Field-Guide for Agentic Engineering — Frank Coyle, UC BerkeleyThe candidate’s compaction, termination inspection, and constrained-agent guidance is reinforced by Codex’s server-side compaction and read-only review boundary.Multiplayer agentic engineering — Arjun Singh, SuperconductorSuperconductor’s isolated cloud sessions establish isolation as a team workflow requirement; Codex adds lower-level sandbox and permission-review mechanisms for enforcing that boundary.
Context Map
agentcodingsecurity#harness-engineering#context-engineering#sandboxing
Uncertainty
At **1,000 tokens per second**, GPT-5.3 Codex Spark made network round trips the bottleneck, motivating a persistent WebSocket mode. These details describe the current harness and API surface; both are expected to shift with later models.