The biggest story this week isn't a flashy new feature. It's three consecutive Claude Code patches in under a week, a signal that Anthropic is stress-testing the boundary between model output and real developer systems at scale. Meanwhile, OpenAI's Codex hit 300 tokens per second, GitHub Copilot quietly gained hands on your mouse, and Cursor added more models to its growing selector. Here's what shipped, ranked by what you actually need to act on.

Claude Code shipped 2.1.289, 2.1.290, and 2.1.291 in rapid succession to fix regressions introduced during the Mods rollout, including lost messages and broken cloud sessions. OpenAI's Codex CLI hit 0.160.0 with multi-agent execution and up to 8x speed gains. GitHub Copilot entered public preview for computer use. If you're running coding agents in production, version pinning is no longer optional.

Anthropic's Claude Code team shipped version 2.1.287 on October 1 with Claude Code Mods, a TypeScript hook system that lets teams rewrite prompts, block or retry tool calls, approve or deny permissions, redact tool output, and modify interface behavior. The 2.1.287 release alone carried 106 changelog entries, including 67 fixes, a scope that foreshadowed instability. What followed was a regression sprint:

Three patches in roughly three days. That cadence tells you something important: the Mods architecture is touching core session and permissions infrastructure, and the blast radius of changes is larger than standard feature additions.

The patches are the noise. The Mods feature is the signal. TypeScript Mods give engineering teams a programmable control plane for Claude Code's behavior, something that was previously impossible without forking or wrapping the tool entirely. What you can do with Mods:

This is governance infrastructure, not a developer convenience feature. Teams running Claude Code in regulated environments or on sensitive codebases now have a legitimate path to customization that doesn't require trusting every model decision unchecked. The catch: Mods are TypeScript hooks, which means they require engineering investment to write, test, and maintain. This isn't a settings toggle. Treat Mod development like you'd treat writing middleware for a production API.

Three patches in one week is not a sign that Claude Code is unreliable. It's a sign that Anthropic is iterating aggressively and fixing fast. The more important question is whether you were pinned to a stable version or riding the latest tag. If your CI pipeline or team default was pointing at the latest Claude Code release, some engineers may have hit sessions that silently dropped messages or cloud sessions that broke mid-task this week. Neither failure mode announces itself loudly. They erode trust quietly. The fix is operational, not philosophical: treat Claude Code like production infrastructure.

OpenAI shipped Codex CLI 0.160.0 on October 1, covering 55 pull requests since 0.159.0. The headline numbers are significant: Codex now generates up to 300 tokens per second, representing up to 8x faster performance in the CLI and up to 6x faster through the API. For agentic workflows, that speed delta is not cosmetic. A coding agent running at 300 tokens per second completes a refactoring pass or test generation loop in seconds rather than minutes. The compounding effect across a team running dozens of parallel agent sessions is meaningful. The 0.160.0 release also advances multi-agent execution and cloud environment support. This positions Codex less as a solo autocomplete tool and more as an orchestration layer: one agent spawning sub-agents, delegating tasks, and assembling outputs. Engineering leaders evaluating Codex should be thinking about how their permission models, secret management, and audit logging hold up when a single user action can fan out into multiple agent executions.

GitHub Copilot entered public preview for computer use on October 2, enabling its desktop applications to click, type, press keys, and scroll in legacy or GUI-only software. This is a qualitatively different capability from code generation. An agent that can operate your GUI tools, not just your editor, is an agent that can interact with software that has no API, no CLI, and no integration hooks. The immediate use case for engineering teams is automating interactions with legacy internal tools, enterprise software with no API surface, or testing workflows that require a human-in-the-loop to navigate a UI. The longer-term implication is that Copilot is expanding its definition of "developer work" to include everything visible on screen, not just what's in a text editor. Computer use capabilities require explicit governance decisions. Who can authorize an agent to click through a UI? What's in scope? What gets logged? Copilot's public preview status means these questions are yours to answer before broad rollout.

Copilot also began retiring four models on October 2: Gemini 3.5 Flash, Gemini 3.6 Flash, Kimi K2.7 Code, and Claude Opus 4.7. This is worth noting if any of your team's workflows or system prompts were tuned around these specific models. Model retirement inside a hosted platform is a distribution risk that's easy to overlook when you're evaluating vendors: you don't control the model lifecycle.

Cursor added Z.ai's GLM 5.3 and GLM 5.3 Flash to its model selector on October 1. This continues Cursor's strategy of positioning itself as a multi-model coding platform rather than a single-model tool. The selector now spans Anthropic, OpenAI, Google, and Chinese frontier models. The strategic implication is portability. Teams using Cursor can switch models without switching tools, which matters when model quality shifts, pricing changes, or a new benchmark leader emerges. The risk is the opposite of Copilot's: instead of the vendor retiring your model, you have so many options that teams fragment onto different models with no shared baseline.

For teams running Claude Code:

For teams evaluating Codex:

For teams using GitHub Copilot:

For teams using Cursor:

Every update above shares a common structure: these tools are adding extensibility and autonomy at the same time. Claude Mods, Codex multi-agent, Copilot computer use, and Cursor's model selector all give teams more control over agent behavior. But more control requires more governance work to exercise that control responsibly. The competitive battle among these tools is no longer primarily about which model writes better Python. It's about which tool gives your team the most operational trust: audit logs, permission scoping, session durability, secret redaction, model substitution rights, and rollback paths. Benchmark scores matter. Reliability at the boundary between model output and your production systems matters more. The engineering teams that will get the most out of this week's releases are the ones that already have an opinion about agent governance. The ones that don't are the ones most likely to be surprised by a silent message loss or an unexpectedly authorized tool call. Set the policy before the agent surprises you.