A clearer view of AI agents.
Google has released Agent Development Kit v2.14.0, graduating execution sandboxes and A2A inter-agent tools to stable status while adding concurrent evaluation runs and closing arbitrary shell command risks.
Read the brief ↗
Google has released Agent Development Kit v2.14.0, graduating execution sandboxes and A2A inter-agent tools to stable status while adding concurrent evaluation runs and closing arbitrary shell command risks.
AI-generated illustration with Google Gemini
OpenAI Codex 0.167.0 introduces a dedicated command center for agent sessions, serialized network approval prompts, and in-transcript patch reviews during approval dialogs.
AI-generated illustration with Google Gemini
Anthropic released Claude Code v2.1.317, introducing a setting to block agent actions when SDK hook callbacks fail, network email disclosure prompts, and fixes for subagent and nested runner sessions.
AI-generated illustration with Google Gemini
Microsoft Agent Framework dotnet-1.27.0 adds a .NET consumer skill, hardens LocalCodeAct process capability checks, enforces strict HTTP redirect handling, and addresses tool approval policies and workflow lifecycle issues.
AI-generated illustration with Google GeminiAnthropic Claude Code v2.1.311 adds idle process controls for self-hosted runners and fixes bugs affecting subagent worktree retention, delayed MCP server connections, memory file staleness, and sandboxed Bash commands.
AI-generated illustration with Google GeminiMicrosoft Agent Framework python-1.24.0 introduces default approval settings for local MCP tools, adds experimental strategies for agent-triggered compaction and skill unloading, and enforces stricter checkpoint deserialization checks.
AI-generated illustration with Google GeminiOpenAI released Codex 0.166.0, ensuring thread forks and resumptions preserve user approval and permission settings alongside reasoning-token counts. The update also adds thread queue resumption, network diagnostics, and safer prompt alternatives after safety blocks.
AI-generated illustration with Google GeminiA study across 202 large enterprises found that only three percent possess automated agent coordination systems, error-driven evaluations, and unified context mechanisms.
AI-generated illustration with Google GeminiAnthropic has donated the Agent Skills standard to the Agentic AI Foundation under the Linux Foundation, shifting oversight of its procedural skill format and assets to a multi-vendor body.
AI-generated illustration with Google GeminiAnthropic updated Claude Code to v2.1.309, closing sandbox path and shell startup escapes, propagating tool restrictions across split-pane agent teammates, and tightening command permission checks.
AI-generated illustration with Google GeminiMicrosoft released python-1.23.0 of its Agent Framework, introducing a breaking migration to Model Context Protocol 2.x foundations, explicit registration scopes for AG-UI endpoints, and tighter execution constraints across shell tools.
AI-generated illustration with Google GeminiGoogle released Agent Development Kit v2.13.0, introducing TypeSafe classification routing, live voice with barge-in support, shared MongoDB sessions, and breaking changes to OpenTelemetry tracking.
AI-generated illustration with Google GeminiMicrosoft Agent Framework dotnet-1.26.0 rejects duplicate tool approval identifiers, preserves toolbox OAuth consent, and adds gRPC selection alongside status message preservation for agent-to-agent interactions.
AI-generated illustration with Google GeminiOpenAI released Codex 0.164.0, introducing enterprise MCP authentication support for confidential OAuth clients, credential masking during remote shell runs, and tightened sandbox security checks.
AI-generated illustration with Google GeminiAnthropic updated Claude Code to version 2.1.307, adding OpenTelemetry tracking for tool calls halted by safety checks, restricting HTTP tool retries to idempotent calls, and improving coordination between parallel fork subagents.
AI-generated illustration with Google GeminiMicrosoft has released python-1.21.0 of its Agent Framework, isolating provider session state when agents run as tools, adding tool-result standing guidance, and serializing concurrent file edits.
AI-generated illustration with Google GeminiGoogle released Agent Development Kit v2.12.0, adding an after-tool callback to inspect tool outputs, human confirmation controls for workflow nodes, and custom service accounts for Agent Engine deployment.
AI-generated illustration with Google Gemini
GitHub published a guide outlining four ways developers can configure the Copilot app to follow codebase rules, execute repeatable procedures through skills, define custom agent roles, and access external tools through Model Context Protocol servers.
AI-generated illustration with Google GeminiAnthropic released Claude Code v2.1.304, introducing explicit person deciders for mod tool checks, guided auto-mode setup across cloud environments, and safeguards against runaway subagents.
AI-generated illustration with Google GeminiGitHub published guidance detailing how developers can open ongoing GitHub Copilot app sessions directly inside VS Code, preserving conversation history, isolated worktrees, and active pull request branches.
AI-generated illustration with Google GeminiOpenAI released OpenAI Agents SDK Python v0.24.0, adding run-wide tool exception handling, RunState schema 1.21 validation, and hardened local execution controls.
AI-generated illustration with Google GeminiAnthropic released Claude Code v2.1.300, introducing mod hooks that rewrite workflow agent prompts and models, OpenTelemetry GenAI trace attributes, and strict input validation for tool hooks.
AI-generated illustration with Google GeminiMicrosoft Agent Framework python-1.22.0 introduces breaking changes that prevent client state from overwriting tool approvals, binds workflow interrupts to resolved thread IDs, adds Python 3.15 support, and alters persisted state formats.
AI-generated illustration with Google GeminiCodex 0.163.0 enables concurrent read-only tool execution, mid-turn steering interrupts, and tighter network proxy validation before DNS lookups.
AI-generated illustration with Google GeminiOpenAI's Codex 0.162 release adds worktree tools and improves approval links, alongside fixes to sandbox behaviour on Linux and Windows.
AI-generated illustration with Google Gemini
A production-agent study examines reviewer differences and changing systems, showing why repeated evaluation needs comparable ratings and current evidence.
AI-generated illustration with Google GeminiA new preprint compares agent-built evaluation suites with established benchmarks and finds that matching visible tests does not guarantee matching hidden ones.
AI-generated illustration with Google GeminiThe Python framework adds controls for stopping runs and requesting tool approval, plus a bounded model-consultation tool and SQLite memory storage.
AI-generated illustration with Google GeminiA GitHub guide shows how developers and agents can work through a shared board or checklist, created and refined with plain-language instructions.
AI-generated illustration with Google GeminiSales Development Agent is in public preview, with autonomous outreach and follow-up. Copilot credit consumption starts on 16 November.
AI-generated illustration with Google GeminiReviewBench tests AI reviewers on 219 pull requests across 19 languages, measuring useful findings as well as missed issues.
AI-generated illustration with Google GeminiThe GPT-6 guide recommends measuring task success, latency and cost, with explicit boundaries for long-running work.
AI-generated illustration with Google GeminiNo stories in this topic yet. Try All stories.
Original briefs with direct source links and publication dates.
Practical guides for using, choosing, building and operating agents. Start with your role, then explore the provider tools.
Understand goals, tools and decisions, with one task explained as a chat response, a workflow and an agent.
4 min readDefine the business result, limit the first task and compare the agent with your current process.
4 min readFollow a task through context, model decisions, tools, permission checks, state and recovery.
4 min readMap identities and data, restrict actions, monitor failures and prepare a tested stop procedure.
4 min read
