AI agents are no longer a side project; this new LangChain report makes it clear they’re becoming serious production systems. Reading it as someone building agents for marketing workflows every day, a few points really stood out.
What the report says
- 57% of organizations already have agents in production, and another ~30% are actively building with plans to ship.
- Customer service and research/data analysis lead as primary use cases, but internal workflow automation is catching up fast.
- Quality is still the number one blocker to production, with latency now the second biggest challenge as agents move into user-facing roles.
How teams are making agents reliable
- 89% of orgs have some form of observability, and over 60% use detailed tracing to inspect each agent step and tool call.
- Just over half run offline evals, while a smaller (but growing) group layer online evals on top of that.
- Most teams are not fine-tuning; they rely on base models, RAG, and better context engineering instead of heavy training pipelines.
The agent stack in practice
- OpenAI still dominates, but 75%+ of teams use multiple models in parallel instead of betting on a single provider.
- A meaningful chunk of companies run open-source models in-house to control cost and meet data or regulatory constraints.
- Coding agents (Cursor, Copilot, Amazon Q, etc.) and research agents are the tools people actually use daily, with many teams layering their own custom LangChain/LangGraph agents on top.
Why this matters for my work
- A lot of my day-to-day at Express Analytics is exactly this: agentic systems for Google Ads automation, glued together with observability, evals, and multi-model decisions.
- Seeing that 32%+ of teams cite “quality” as the main blocker resonates with what I’m optimizing for: fewer hallucinations, better guardrails, and faster feedback loops on real campaigns.
If you’re building agents too
- This report is a good reality check: agents are past the hype stage, but the hard problems now are debugging, measuring, and improving them.
- If you’re working on agents (or want to), happy to chat about observability setups, eval workflows, or how we’re using agents for marketing analytics in production.