Agent Zero v2.7: Reliable Responses, Cerebras, & Browser Proxies

v2.7 makes response handling more dependable, adds Cerebras and project-scoped Browser proxies, verifies remote Linux computer-use targets, and fixes prompt, parallel-call, and Codex OAuth behavior.

Agent Zero
#releases

Agent Zero v2.7 changelog banner

🔁 Response & Streaming Reliability

  • Native responses follow the tool path — Responses output text now runs through the normalized response-tool executor, preserving goal overrides and state cleanup while rejecting empty payloads.
  • Partial streams stay readable — Incomplete tool arguments no longer fail during streaming, text and legacy message fields both render, and live logs remain synchronized with partial snapshots.

🧩 Prompt & Tool Routing

  • Prompt JSON reaches the required wire format — JSON fence markers are stripped from bundled prompt examples at render time while non-JSON fences remain intact.
  • Final responses stay top-level — Response calls are kept outside parallel groups so they correctly end the parent message loop without unnecessary follow-up turns.

🧠 Providers & OAuth

  • Cerebras provider added — Cerebras is now available through native Chat Completions support with onboarding guidance, a production default model, and a bundled logo.
  • Codex reasoning effort normalized — Legacy reasoning_effort values are converted at the Responses proxy boundary while native reasoning objects remain unchanged.

🌐 Browser & Computer Use

  • Project-scoped Browser proxies — The internal Docker Browser supports HTTP and SOCKS proxies, bypass rules, and optional authentication through Playwright, with grouped settings and automatic runtime restart.
  • Remote Linux targeting hardened — Agents verify window focus before target-bound typing, use scoped snapshots and observed receipts, and stop repeating identical failed actions.