CODEX 0.120 ADDS BACKGROUND AGENT STREAMING; GPT‑5.4 PITCHED FOR END‑TO‑END CODING AMID MIXED MODEL FEEDBACK
OpenAI shipped Codex updates for agents and tooling while positioning GPT‑5.4 for real multi‑step coding work, but some users report reasoning regressions. Per...
OpenAI shipped Codex updates for agents and tooling while positioning GPT‑5.4 for real multi‑step coding work, but some users report reasoning regressions.
Per the Codex 0.120.0 notes, Realtime V2 can now stream background agent progress, TUI hook visibility is clearer, MCP tool typing is stricter via outputSchema, and Windows sandbox/TLS issues got fixed release notes. That’s a solid quality‑of‑life bump for anyone running agentic workflows in production.
An independent write‑up details how OpenAI is framing GPT‑5.4 as a coding+reasoning+agent system designed for multi‑step work across code, documents, and tools rather than simple autocomplete analysis.
At the same time, community threads flag mixed results: some prefer 5.1 Thinking over 5.2, and others see the 5.4 Thinking model reopening already‑settled points without identifying concrete gaps (thread 1, thread 2).
Background streaming and tighter tool typing reduce blind spots and flaky integrations in long‑running agent tasks.
Model behavior appears in flux, so pinning versions and validating reasoning quality remains essential.
-
terminal
Enable Realtime V2 background streaming and measure stuck‑work detection, tail latency, and operator triage time in your agent pipelines.
-
terminal
Adopt MCP outputSchema typing for tools and assert JSON contracts end‑to‑end; compare failure rates before and after.
Legacy codebase integration strategies...
- 01.
Pin current model versions, add health checks, and keep fallbacks if 5.4 Thinking regresses on your tasks.
- 02.
If you run Windows hosts or remote TLS connections, re‑test sandboxes and websocket flows after the Codex fixes.
Fresh architecture paradigms...
- 01.
Design agents as long‑running workers with background progress channels and typed tools from day one.
- 02.
Model orchestration should treat reasoning and tool use as a single loop; bake in retries, state, and observability.