Agents get a voice, a memory, and a tighter grip on credentials

Agents learn to talk

The week's most visible shift is that coding agents started speaking. Codex CLI 0.155.0 added experimental /voice conversations with live transcripts and microphone controls, enabled through /experimental, and the same release brought Touch ID verification for MCP requests in local TUI sessions on supported Macs — a neat pairing of "let it listen to you" and "make sure it is you." Devin shipped a competing take a day earlier with live voice mode, moving calls onto a live speech model with ringtone-while-connecting, hold-Space-to-talk after muting, ordered spoken replies in the transcript, and Slack thread attribution showing who was on the call. Codex also polished the non-voice TUI, adding live reasoning summaries and completion timestamps, though 0.155.1 walked the reasoning summaries back to disabled-by-default in new local sessions after providers without support rejected the requests — a reminder that these surfaces are still settling.

Codex for iOS picked up folders in the file picker and writing blocks with draft alternatives, plus fixes for queued prompts vanishing when you reopen a task.

Claude Code hardens the edges

Claude Code's week was less about features and more about closing gaps that had been quietly costing users. v2.1.275 fixed a restored memory file's age note changing between requests after compaction or resume — the kind of small nondeterminism that causes prompt cache misses and inflated bills — and stopped plugin and marketplace messages, logs, and listings from leaking passwords or tokens stored in git, ssh, or marketplace URLs. It also added a send-now key (ctrl+enter) that interrupts the current turn and flushes queued messages at once, and syncing of claude.ai skills and plugins into terminal sessions.

2.1.274 tackled two failure modes that previously left users stuck: sessions endlessly retrying unexpected tool_use_id 400 errors now self-heal corrupted transcripts where possible and otherwise exit with a clear error and a /rewind hint, and a new CLAUDE_CODE_MCP_STARTUP_WAIT_MS variable bounds how long the first non-interactive turn waits on connecting MCP servers.

Elsewhere the sequence reads like a bug-fix relay. 2.1.276 patched a 2.1.275 regression where every request through a proxy or gateway failed with 400 ... Input tag 'advisor_20260301' — a one-line reminder of how fast a release train can break a setup. 2.1.273 added opt-in gateway hint headers and reverted a change that applied Read and Edit deny rules to Bash lines the permission checker can't analyze, so time -p make build prompts again instead of being denied outright.

Coarser controls, better audits

Two releases reworked where policy actually lives. v2.1.278 made auto mode default to the server-side classifier for the Claude API, Bedrock, Vertex, Foundry, and gateways, with CLAUDE_CODE_AUTO_MODE_SERVER=0 to opt out and a new row in /status confirming where the classifier runs — meaningful because the server-side path doesn't bill classifier overhead. 2.1.271 brought fast mode to Remote sessions on cloud and self-hosted runners and added per-command allowed_domains so each Bash, PowerShell, and Monitor command reaches only the hosts it needs. v2.1.277 filled an ecosystem gap by falling back to AGENTS.md when no CLAUDE.md exists and fixed claude -p and Agent SDK sessions hanging with no result after an internal error.

Devin's September 18 release is the one to read before touching your integrations: the Enterprise API audit-logs list response no longer populates total, returning null, so pagination must use has_next_page and end_cursor. It also added Origin session filters, 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 60 fps screen recordings. MCP credentials moved to each server's page, part of the same week-long theme: agents are getting more capable and, finally, more careful about the keys they hold.