Agent reliability & evals

CrewAI 1.15.23 lets your agent pipelines be scored in production and survive rate limits — tracing-based evaluation scores each run, and throttled model calls now retry automatically. If agents carry campaign work, upgrade in staging first.

Your LLM observability will break if you upgrade LiteLLM this week — the new release candidate migrates its Langfuse tracing callback to SDK v4, forcing anyone using Langfuse to rewrite integration code before deploying.

Anthropic released a new model, claude-opus-5-5, callable directly from its Python SDK — if you route marketing tasks across different AI models, re-run your tests before switching anything in production.

Agents can now work from your company's scattered files and cite their sources — V7's new platform turns internal docs into agent-ready memory, replacing workflows that today depend on someone digging through shared drives.

OpenAI is opening an ad surface inside agentic experiences — Sponsored Agents lets you run paid campaigns inside agent answers, managed through HubSpot and Shopify. That is a new paid channel every marketing engineer must evaluate now.

promptfoo shipped version 0.123.0 — the open-source eval harness now covers the latest frontier models, but a default switch for newer GPT models breaks some existing test configs. If you run model evals, test before upgrading.

n8n, the workflow automation tool many marketing teams use to wire LLMs into campaigns, shipped version 2.39.0 with fixes to its AI agent nodes and security hardening across webhooks and secrets. If you run production marketing automations through n8n, this is a release worth testing before the next campaign cycle.

n8n shipped three releases in quick succession, fixing the reliability and integration issues that break AI agent workflows in production — Anthropic thread recovery, task runner timeouts, webhook handling, and secrets failover. If you run marketing automation through n8n, this week's upgrade is a maintenance task with real payoff.

Google shipped stateless updates to MCP that let agents scale horizontally without sticky session state — if you're pushing marketing agents toward production, this removes a core infrastructure bottleneck that has made concurrent personalization and lifecycle workflows hard to parallelize.

Google DeepMind shipped a faster Gemini Flash variant and agentic video understanding — two releases that could change how you route model calls and automate creative QA. PostHog open-sourced agent skill templates and Langfuse added eval alerting for production LLM workflows.

Three platform ships this week lower the barrier for deploying AI agents in marketing workflows — Google's MCP stateless scaling, Vercel's Claude Managed Agents in Chat SDK, and Vercel's dashboard agent deployment. But Twilio's data showing 78% of consumers actively bypass AI agents is a reminder that infrastructure readiness and consumer willingness are different problems.

Databricks introduced a database architecture pattern for high-throughput agent workflows on Postgres — if your real-time personalization depends on agent loops, traditional transactional databases may be your throughput bottleneck. Three other releases this week touch LLM eval workflows, marketing-to-deployment pipelines, and HubSpot's platform direction.

Langfuse shipped two releases this week upgrading LLM evaluation workflows and agent observability. If you trace marketing agent calls or run evals on campaign prompts, reusable eval filters, execution SLO metrics, and OpenAI-compatible agent support are the headline features worth pulling into your stack.

Langfuse shipped a rebuilt eval experience, LangChain patched a Claude agent crash mode, and Supabase plus Bedrock added infrastructure for governed, context-aware agents — all relevant if you ship AI in marketing systems.

Langfuse shipped agent turn tracing for multi-step approval workflows, OpenAI's SDK added Bedrock and shell streaming support, and a compact multilingual embedding model landed — three builder tools worth evaluating this week.

Vercel shipped a one-command setup for coding agents that routes through its AI Gateway — if your team experiments with AI-assisted content workflows, this removes the infrastructure configuration overhead.

Vercel shipped a one-command setup for coding agents that routes through its AI Gateway — meaning marketing ops teams can spin up AI-assisted landing page builds and campaign experiments without wrestling with complex config.

Langfuse shipped per-conversation tool approval for in-app agents — if you're building agent workflows, this reduces friction by letting users approve a tool once per conversation instead of every call. Also relevant: n8n hardened AI agent execution with better schema handling and leak prevention.

Google shipped a production-ready agent evaluation service in Gemini Enterprise this week — the first turnkey option for teams that need to measure agent quality without building custom testing infrastructure from scratch.