Model selection and tool mode participate in turn construction rather than living only in the visible prompt.
Agent Atlas / Coding agents
openai/codex
A local coding agent whose public Rust runtime makes session state, sandbox policy, tool modes, subagents, and operator control unusually inspectable.
Coding agent
Full dossier
Reported + sourced
Reviewed July 21, 2026
Intended operator
Developers working from a terminal, editor, app, or an integrating client.
Primary surfaces
CLI and TUI / app-server protocol / editor and application clients
Tracked repository
Teaching mechanism
The visible prompt is only one input to the first request
The runtime can inspect pending input, resolve permissions and tool mode, run hooks, and add higher-authority context before a model invocation is assembled.
Evidence limitAn interception point does not prove that injected instructions are correct, disclosed, or consistently enforced.
Inspect the source
System Map
Eight inspection boundariesThread history, goals, skills, instructions, and rollout metadata are assembled and persisted through separate runtime paths.
Typed tools are planned by the model and executed through runtime handlers rather than by the model itself.
Permission profiles and sandbox policy travel with session and turn state.
Thread stores, rollout records, interruption handling, and resume views make work durable and reconstructable.
Subagent lineage, startup hooks, lifecycle events, and send/receive receipts distinguish parent and child work.
The operator can review, interrupt, approve, redirect, and resume work through client surfaces.
Structured events, diffs, tool results, thread metadata, and public tests provide inspectable receipts.
Learn Through This Project
8 lessons
Chapter 1
Before the First Model Request
Chapter 2
The Invisible Prompt
Chapter 3
Tool Calls: Contract, Authorization, Execution
Chapter 4
Loops, Workflows, and Recovery
Chapter 5
Delegation Is a Distributed System
Chapter 7
The Approval Trap
Chapter 8
The Human Control Loop
Chapter 9
Trust Must Make Its Case
Latest Reporting
8 articles
The Diff Can't Be Guesswork Once an Agent Writes
July 20, 2026 — OpenClaw is making a simple rule explicit: when an agent writes a file, a review surface should show the patch it can prove—and show uncertainty when it cannot. openclaw/openclaw, openai/codex
Codex Stops Asking Its Subagents the Same Question TwiceJuly 20, 2026 — A Codex TUI patch cuts duplicate status reads by treating new, resumed, and actively streaming agent threads as different lifecycle states. openai/codex, NousResearch/hermes-agent
Codex Keeps the Prompt After You Interrupt ItJuly 16, 2026 — A new Codex TUI change treats an interrupted request as part of the record, then gives the next instruction a clean composer. openai/codex
Open Interpreter: roots, maintainers, and the contributor advantageJuly 16, 2026 — The new Rust Open Interpreter is openly building on Codex; its public advantage is upstream leverage, not a proven inside track. openinterpreter/openinterpreter, openai/codex
Codex Makes Thread History a Rebuildable ViewJuly 14, 2026 — Two fresh commits separate a coding agent’s durable JSONL record from its fast SQLite history view—and teach that view how to catch up after it falls behind. openai/codex
Agent Runtimes Move Trust To The Exit GateJuly 9, 2026 — OpenClaw, Codex, and Hermes all changed runtime trust surfaces this week, making the last moment before a tool call, secret handoff, or approval the place operators need to inspect. openai/codex, openclaw/openclaw, NousResearch/hermes-agent
Builder Work: Ralphy Gives Coding Agents A Video RuntimeJuly 9, 2026 — Ralphy is today's Builder Work Highlight because Alec's repo turns a coding agent into the operator of a reproducible video pipeline, not just a prompt wrapper around media APIs. openai/codex
Harness Engineering Is Becoming The Field's Self-Improvement LayerJuly 8, 2026 — Lilian Weng's July 4 map of self-improving harnesses gives builders a constructive middle path: improve the runtime around the model, but keep evidence, permissions, and evaluation legible enough that the loop can be trusted. lilianweng/lilianweng.github.io, openai/codex, different-ai/openwork, openclaw/openclaw, NousResearch/hermes-agent, earendil-works/pi, anomalyco/opencode
Primary Sources
Open Questions
- Which effective instructions and permissions can every client show before a turn begins?
Evidence state describes what this Atlas profile can support today. It is not a product score and it is not evidence that uninspected capabilities are absent.