Coding agents learn to talk, and to clean up after themselves

Agents find their voice

The most visible shift this week was agents learning to hold a conversation. Codex CLI's experimental /voice mode brings live transcripts and microphone controls to supported builds, gated behind /experimental, alongside Touch ID verification for MCP requests in local TUI sessions on supported Macs and live reasoning summaries in the status row. Devin went further and put voice calls on a live speech model, with proper call controls, hold-to-talk after muting, ordered spoken replies in the transcript, and Slack thread attribution showing who joined a call and for how long. Devin also started waking sleeping sessions as you type, and made its Review feature always check for security bugs rather than leaving it behind a toggle.

Codex's voice debut came with a same-week caveat: reasoning summaries are now off by default in new local TUI sessions, fixing rejections from providers that don't support them, while explicit settings are still honored. On iOS, Codex gained folder creation in the file picker plus writing blocks with draft alternatives, and fixed queued prompts that went missing or looked stale after reopening a task.

Claude Code's week of self-repair

Claude Code's eight releases read mostly as one long reliability push. The headline fix: sessions stuck endlessly retrying unexpected tool_use_id 400 errors now self-heal corrupted transcripts, or end the loop with a clear error and a /rewind hint, with a new env var bounding how long a first non-interactive turn waits on connecting MCP servers. A restored memory file whose age note changed between requests was causing prompt cache misses after compaction or resume — fixed, along with plugin and marketplace messages no longer leaking passwords or tokens stored in git, ssh, or marketplace URLs. Projects without a CLAUDE.md now fall back to AGENTS.md, and claude -p and Agent SDK sessions no longer hang silently after an internal error.

Gateway and proxy users got several turns of attention. Claude Code added opt-in hint headers for LLM gateways and reverted a Bash deny-rule change that had been blocking commands the permission checker couldn't analyze, then patched a regression that broke every request through a proxy or gateway. Auto mode now defaults to the server-side classifier on the Claude API, Bedrock, Vertex, Foundry, and gateways, avoiding classifier overhead charges and surfacing which classifier a session uses in /status. Remote sessions gained fast mode on cloud and self-hosted runners, along with per-command allowed domains for sandboxed Bash, PowerShell, and Monitor commands.

Enterprise plumbing

Devin's enterprise surfaces got their own pass: session origin filters for Slack, Web, API, Jira, and Linear; an MCP reconnect action that redoes authentication without uninstalling; per-PR Devin Review from the Cmd+K palette; Azure DevOps Server 2020/2022 connections; and new ActiveCampaign and Grafana MCP integrations. Note for API consumers: the audit-logs list response no longer populates total — paginate with has_next_page and end_cursor.

Grok Build's quieter week still matters for anyone relying on it: Memory is now generally available, and earlier in the week it fixed macOS image paste and session history loss, stopped cancelled MCP tool calls from continuing in the background, and added structured JSON to MCP tool results.