← Field Notes
NOV 18 · Release · via OpenAI Engineering Budget ScopingPermissionsReach Preview

OpenAI explains how Codex limits what its coding agent can touch

The boundary is set before the agent moves, not negotiated as it goes. That shifts the design question from 'did you approve this action' to 'did you understand what you were granting when you started.'

FIG. 1 · Where this sits in Budget Scoping
    18 NOTES ON BUDGET SCOPING IN 26 WEEKS · SCRUB THE WEEKS, CLICK ONE TO PIN IT
    Published, annotated field notes attached to Budget Scoping, by week observed · last 26 weeks · n = 18. Rail markers are lifecycle events from the pattern changelog.
    Data table
    Week ofNotes
    AUG 24AID-Guard counts what an agent already did before it allows more · Martin Fowler counts the context each agent handoff has to rebuild
    AUG 31Spotify's Portal cut what Claude Code reads by 90 percent
    SEP 7Seven agents ran real businesses and sent $12,431 in unasked invoices
    SEP 14Suganthan watched Claude pay a penny to read his website · AcquireBound makes agents ask before grabbing new access
    SEP 211,248 GitHub agent specs show safety steps missing from most · 73 000 OpenClaw posts show humans shape what an agent may do
    OCT 5Vercel lets agents buy domains but makes them stop for your yes · arXiv paper maps how humans govern thousands of research agents
    OCT 12EAIR paper finds enterprise agent teams break rules together · Amex lets cardholders cap what an AI agent can spend before it shops · Fowler and Jain measure what scaffolding costs to run an AI coder
    OCT 19Researchers find agent pricing follows whatever prices you show it
    OCT 26Tracekit Box locks agent actions behind signed, fuzz-tested approvals
    NOV 2CLEO cuts the model size 100x and still passes 83.8% of tests
    NOV 9Grok Build's coding agent exposed confidential SpaceX data by default
    NOV 16OpenAI explains how Codex limits what its coding agent can touch

    Machine summary of the source

    OpenAI published an engineering explainer on how its Codex coding agent restricts what it can read, write, and run depending on the platform it runs on. The agent works inside a constrained environment. That environment controls which files the agent can open, which commands it can run, and whether it can reach the network. The designer decides those limits before handing work over, not by approving each action as the agent runs. The limits define the trust relationship. A narrow scope, say read-only file access with no outbound network, means the human is granting the agent a small, legible reach. A broader scope means the human accepts more risk in exchange for fewer interruptions. OpenAI frames this as a deliberate trade-off. Tighter limits make the hand-off safer. They also mean the agent asks for help more often. For a product team, granting access is a decision the human makes before the agent starts. The human needs to see what access they granted. They need to understand what that access allows. They need to be able to narrow or revoke it. Without that view, the hand-off is opaque and trust is guesswork.

    The summary above is generated; the note at the top is the editorial judgment. Primary source ↗