This Week in Agentic Coding #3

TLDR

Anthropic moved a lot of agent API surface out of beta: computer use, Files API, Agent Skills, Enterprise user management, plus the new browser use tool.
Cursor is building toward always-on coding agents with PR and Slack subscriptions, scheduled tasks, /goal, custom modes, and isolated subagent machines.
Claude Code added the boring controls serious teams ask for: loop usage stats, curated model pickers, prompt-cache TTLs, organization pricing, and keyless Console login.
Cline focused on durable agent operations: smaller hub event logs, replay after restarts, MCP marketplace enforcement, and a fixed global MCP toggle.

Models & APIs

Claude's agent APIs are leaving beta in a bundle (release notes)

Anthropic made computer use generally available on the Claude API as computer_toolset_20260801, launched browser use as browser_toolset_20260801, and moved the Files API, Agent Skills, the Skills API, and Claude Enterprise user-management Admin API endpoints out of beta. The browser tool is the bigger product clue: it reads a hosted browser viewport through the accessibility tree, elements, forms, and tabs, then adds form input, tab management, download reporting, and opt-in file upload. If you maintain a Claude integration, this is a migration week: remove stale beta headers where the new surface is stable, then audit request shapes before upgrading computer use. August 19.

OpenAI gives Kiro a spec-driven GPT-5.6 path (announcement)

OpenAI says the GPT-5.6 family is now available in Kiro, with Sol, Terra, and Luna inside workflows for planning, building, reviewing, and testing software. The source frames Kiro's value around structured requirements, technical designs, executable tasks, codebase context, team standards, checkpoint review, and property-based testing. The only number worth repeating is the one OpenAI published: on Terminal-Bench 2.1, testing found GPT-5.6 Terra completed successful tasks in Kiro at roughly 82% cost reduction. Treat the number as vendor-provided, then steal the operating lesson anyway: better task structure beats bigger prompts. August 24.

OpenAI tries to keep ZDR viable for frontier agents (preview)

OpenAI previewed Private Safety Processing for eligible Zero Data Retention API deployments. The promise is narrow: OpenAI says customer prompts and model responses are not retained after processing, customer content is not available to OpenAI personnel for review, and enterprise customer data is not used for training unless the customer opts in. Private Safety Processing is meant to detect risk patterns across related interactions while returning limited safety signals instead of exposing the underlying content. This matters for agent builders because the hardest safety failures show up across a task, after a single-turn scanner has already said yes. A technical white paper is planned for September. August 19.

Agentic Coding Tools

Cursor agents now wake from events and hold goals (changelog)

Cursor added cloud-agent subscriptions for PRs, Slack threads, and scheduled tasks. Cursor says cloud agents automatically subscribe to PRs they create, then drive them to completion by fixing CI and addressing bot comments. The same release added custom modes, which pin a skill as the active mode in chat, subagents on separate virtual machines with isolated project copies, /goal for long-lived objectives, and steering messages that wait for the next tool call instead of interrupting the agent mid-action. This is the right direction: agent products stop being demos when they can remember what they own and wake up at the correct time. August 19.

Claude Code exposes the knobs teams need before autonomy scales (changelog)

Claude Code versions 2.1.234 through 2.1.245 shipped a dense runtime-control week. The most useful additions are a Loops breakdown in /usage, modelPicker for curating /model, promptCacheTtl and subagentPromptCacheTtl, managed modelPricing for contracted rates in /cost, keyless Anthropic Console sign-in, and model plus effort visibility for subagents in /tasks. The fixes are just as practical: remote MCP reconnection in non-interactive and SDK sessions, background subagent wakeups after Bash completes, a timeout when the API never starts responding, proxy fixes, and a glibc 2.44 startup crash fix. If your team runs agents all day, this is a reduce-operational-weirdness update.

Cline makes long-running hubs less fragile (SDK release)

Cline capped the hub's durable event log at 64 MiB on disk after event envelopes carrying full session snapshots could grow it into tens of gigabytes. Oldest events are dropped first, the database is vacuumed to return space, and pruning also runs after every 16 MiB appended. The same SDK release fixed task.completed telemetry being dropped for most interactive sessions and refreshed the model catalog. Pair that with the prior restart work in this week's releases and the theme is clear: the hub is becoming something you can leave running without babysitting the SQLite file.

Gemini CLI keeps investing in failure visibility (release)

Gemini CLI v0.57.0-preview.0 added eval validation, a tool-call formatter, and integrated failure summaries. The release also fixed Cloud Workstations OAuth redirect URI resolution, IDE directory mismatch handling, silent retries and availability TTLs for capacity errors, full rollback of a multi-turn request on cancellation or abort, and workspace state mismatches in git environment handling. None of that is flashy. All of it is the kind of engineering that turns agent runs from mysterious transcripts into inspectable systems. August 19.

MCP Ecosystem

Chrome DevTools MCP turns browser agents into debuggers (directory)

Chrome DevTools MCP gives assistants direct browser control through Chrome DevTools, with Puppeteer-based interaction, network analysis, console debugging, screenshots, accessibility-tree snapshots, and performance tracing with Chrome DevTools insights. That is a stronger surface than screenshot-and-click automation because the agent can inspect the same failure evidence a frontend engineer would inspect. The directory diff marked it as a top MCP mover this week with +2,141 stars and +7,613,913 downloads.

📊 New in the Directory

find-skills (skill): discovers installable agent skills when a user needs a specialized workflow. It led the skills movers in the diff with +344,512 downloads, which says the skill layer now needs its own discovery layer.

Mnemo (MCP server): gives agents persistent memory with hybrid full-text and semantic search, local embeddings, SQLite-vec storage, and cross-machine sync. New in the MCP snapshot with 81,449 downloads.

Vaara (MCP server): records runtime evidence for agent tool calls, scores risk, and writes hash-chained audit records for AI Act compliance. New in the MCP snapshot with 76,719 downloads.

claude-code-plugins (plugin): packages official Claude Code plugins for Agent SDK development tools, PR review, and commit workflows. New in the plugin snapshot with 138,799 stars.

Browse all at getagentictools.com.

Be pragmatic, not dogmatic.

Someone forwarded you this? Get The Agentic Coding Playbook, free.

Want to reach thousands of developers building with AI? Sponsor the digest.