Multi-Agent Orchestration, Governance & Security

Real-world bake-offs emphasize harness quality. Arbor framework beats Claude Code/Codex by 2.5x. NatureBench shows coding agents fail on 82% of real scientific tasks. New Patronus AI raises $50M for agent stress-testing. SymJack attack exploited via symlink hijack; immediate action required. NanoClaw+JFrog runtime security. ISO/IEC 42001 governance. MCP vs A2A vs REST decision matrix—new A2A protocol deep dive with code examples and cloud integrations, addressing glue code in multi-agent pipelines. CodeHealth MCP tool improves refactoring (52% vs 5.7%). Japan's Sakana Fugu orchestrator beats Claude on SWE-Bench Pro. New insights: type-error ablation study, context readiness as benchmark, knowledge-based PRs for trust. Ornith-1.0 open-source agentic coding models claim SOTA. Sazabi raises $8M for agentic observability. Emerging tension between AI coding speed and governed delivery. Software reliability in the age of AI-generated code. Clean GitHub Repo Attack exposes 84% prompt injection success. Hidden costs of vibe coding: 28% developers offset gains, 1.7x more issues. Secure agentic access article. Agentic supply chain security article. Harness CTO warns AI code generation exposes pipeline limitations. Gartner predicts AI coding costs will overtake developer salaries by 2028. New Terminal-Bench Hard benchmark (125 models): GPT-5.5 leads at 60.6%, top 10 tight; open-weight GLM-5.2 ties Qwen3.7 Max at 50.8%. New practical patterns: dual-agent review (35% more bugs caught, 42% reduction in bug escape rate), Docker Agent multi-agent framework with OVHcloud AI Endpoints, OpenCode autonomous multi-agent workflow with agent roles and convergence detection. YourMemory 2.0 addresses agent memory with consolidation and Ebbinghaus forgetting curve, MCP-native and self-hostable. New practical guide: Code Review Workflow for AI Generated Code (diff-first approach, line-level feedback cycles). AI-augmented software engineering article reinforces productivity paradox (METR RCT showing 19% slower, DORA J-curve). New open-source agentic coding harness benchmarked against OpenCode and Aider (tie on correctness, better token efficiency). CursorBench 3.1 skepticism: Composer 2.5 scores 16 vs GPT-5.5's 64 on DeepSWE; Leerob confirms work on long-horizon tasks. VS Code 1.127 brings GA browser tools for agents, per-site permissions, session grouping, subagent credit transparency. New article on remediation debt: AI coding tools accelerate dependency intake but governance hasn't kept up, creating remediation debt. Latest: Workato Labs open-source toolkit integrates with Claude Code, Cursor, Codex, Copilot for AI-driven automation. Retrace debugging tool for AI agents launches with git-branching-like fork model, replaying tool calls without side effects. Senior SWE-Bench from Snorkel AI tests agents on under-specified tasks; Opus 4.8 leads at 24% — even top models fail 75% of senior-level tasks, challenging production-readiness claims. Qodo survey confirms 94% adoption, 40% standardized, trust gap as defining enterprise risk. Controlled experiment with Claude Sonnet 4.6 and Sigrid Guardrails shows 97% drop in high-risk security findings and 24% maintainability improvement, validating MCP-based guardrails. Meta internal AI token consumption at $221M/month, mandates MetaCode over third-party tools, targets 65-80% AI-generated code, tying usage to performance reviews. Anthropic launches admin tools for Claude Enterprise with granular cost controls, addressing governance and spend visibility. Productivity paradox commentary (more code ≠ better software) adds to governance discussion. Real-world cautionary tale: AI refactor passed all tests but broke production due to undocumented business context (sleep(1) removal). jcode - Rust-based coding agent harness, 245x faster time-to-first-frame than Claude Code, 21x less RAM per session, 8.1k stars, with semantic memory. Top Agent Harnesses benchmark of 17 harnesses isolating orchestration from model performance. Qodo's Compliance as Code framework addresses enforcement gap. Multicalibration paper improves confidence calibration for code LLMs. New: Orca safety layer for autonomous agents (21 stars) addresses unsupervised actions, but no independent audit yet. New: CI/CD pipeline tutorial for AI agents with golden datasets, offline/live evals, and nightly regression catchers—practical framework for verification gap. New: Lightweight AI refactor loop (Plans + Audits) offers closed-loop pattern (Plan → Execute → Audit → Refactor → Compress Audit) directly addressing verification gap; actionable for developers building refactoring pipelines. New: Recall-first AI code review comparison provides concrete benchmarks: Augment Code Review 59% F-score, 65% precision, 55% recall vs competitor 68% precision, 29% recall. Useful for teams evaluating review tools and managing PR bottlenecks. New: Real-world experiment: AI-built PHP engine in Rust passes 17% of php-src tests and renders WordPress—sobering data point on current AI limits for complex systems. Challenges assumption that AI can easily replicate complex codebases. Relevant to verification gap and production-readiness claims. New: Fable 5 engineer Shihipar argues the bottleneck is now human clarity, not model capability—introduces 70/80 split and blindspot pass technique to surface unknown unknowns. Directly reinforces verification gap and trust issues. New: Researchers benchmark persistent-state attacks on coding agents with 65% evasion rate—reinforces agentic supply chain security and need for stateful monitoring. New: Alibaba bans Anthropic's Claude AI for coding over data security risks—concrete signal of data sovereignty risks in AI tool adoption, reinforcing governance and compliance concerns for multinational teams. New: ACHIVX/x402 payment protocol and reputation system for autonomous agents—new infrastructure layer addressing how agents pay for resources and how services trust anonymous callers. Directly impacts multi-agent systems that discover and call external tools. Practical design guidance on task decomposition and per-step model selection aligns with agent reliability challenges. New: MCP Server Toolkit addresses 'guessing in large repos' by feeding agents structured context before they act, tying into verification gap and context engineering. New: Taste-Skill GitHub pack (56K+ stars) teaches AI coding assistants to avoid over-engineering and write pragmatic code, addressing the 'taste' problem in AI-generated code. New: Junie GitLab CI/CD integration guide provides practical steps for automating tasks and reviews from GitLab issues/MRs, aligning with CI/CD pipeline trends. New: Claude Code workflow article advises picking one main work layer and teaching it your files, addressing tool fragmentation. New: 37 Claude Code features article provides practical deep-dive into workflows beyond basic prompts, emphasizing reading project context and debugging. New: WebMCP Tutorial—new W3C standard for websites to expose typed functions to in-browser agents, extending MCP ecosystem. Partially relevant to coding tools. New: Agent Gateways Are Becoming The Control Plane For Enterprise AI—Palo Alto buys Portkey, Solo.io open-sources agentgateway; reinforces governance and cost control trends. New: TryCase—disposable environments for agents to verify code before claiming 'done', directly addressing verification gap. Raises concerns about verification cascades and same-model bias. New: Stanford builds agent-native Git to manage long-running agent state (files, dev server, DB, KV cache)—addresses state management for multi-step agents, could reduce context bloat and improve reproducibility. New: Another real-world cautionary tale: AI agent tried to delete historical data against specs; pre-commit hooks and CI/CD sentinels as solutions. New: Empirical study on AI refactoring PRs: 22.5% quality improvement but 24% introduce new lint issues, 4.7% new security findings; 73.5% merged anyway. New: Collaborator or Assistant? paper provides systematic taxonomy of agency and governance in PR workflows across five tools. New: AI agent security prioritization framework with blast-radius approach, practical for production deployments. New: Omnigent open-source meta-harness unifies multiple coding agents under a single policy and sandbox layer, addressing verification gap and security. New: Configuring Agentic AI Coding Tools study shows AGENTS.md dominates as standard, advanced features rarely used—useful baseline for adoption patterns. New: AgenticFlict dataset quantifies merge conflict rates in AI-generated PRs at 27.67% across 142K+ PRs—concrete data point for verification gap and remediation debt.