Enterprise Agents Stop Demoing and Start Doing Measured Work

A ticket agent with a human in the loop

Vestas says it has deployed a Jira Ticket Agent built on Dataiku and Snowflake to support its wind turbine software environment, analysing new tickets to recommend the most relevant component or team. Vestas processes 100 to 150 support tickets per month, and the previous ticket assignment process had a 72-hour service level agreement, while the new target for ticket assignment is less than one hour. Vestas targets a 95 to 99 percent reduction in ticket assignment time and projects up to a 40 percent reduction in overall ticket handling and lead time, though those are targets and projections reported by the company, not measured results from a completed rollout. m:…3784

The deployment detail that matters most is the human-in-the-loop design. The agent currently posts recommendations as comments in Jira rather than automatically changing fields, and Vestas employs a human-in-the-loop approach for ticket triage, with engineers reviewing recommendations before final action. The project combines machine learning, large language models, and retrieval-based search, and Vestas is exploring the integration of turbine logs into the support workflow. That combination of a bounded job, a measurable target, and a human checkpoint is a pattern worth noting: the agent is not being asked to run support, but to route it faster while a person stays in control of the decision. m:…3784

Coding agents meet governed pipelines

Dataiku says it has launched Dataiku Headless, an integration that allows developers to build and manage Dataiku projects directly from coding agents. Dataiku Headless allows developers to use coding agents to build and modify Dataiku projects, uses Cobuild to generate data pipelines, models, and dashboards, and is available as a plugin for Claude Code, Codex, Cursor, Snowflake CoCo, AWS Kiro, and OpenCode. It can be connected to coding agents using an MCP server and skills, requires Cobuild to be enabled on the Dataiku instance for building capabilities, and is designed to ensure that work generated by coding agents is visual and inspectable. m:…0125

Dataiku says Headless also supports migration of legacy workflows from SAS, Alteryx, and Excel into Dataiku, includes administrative tools for investigating and managing Dataiku environments, and allows administrators to find failing jobs within a project. Dataiku Headless respects the permissions of the Dataiku API key used to connect, which is the governance detail that makes the bridge usable in an enterprise setting: the coding agent inherits the permissions of the key, rather than gaining a separate, broader access path. The pitch is not that coding agents replace governed data work, but that governed data work can be built from the coding agent a developer already uses, with the resulting pipeline still visible and reviewable in Dataiku. m:…0125

Benchmarks on real work

Snowflake says it introduced SnowPR, a benchmark designed to evaluate coding agents against its large Java and C++ monorepo, using 65 merged pull requests across 11 product areas that are graded using hidden tests that shipped with the original pull requests. SnowPR tasks are significantly larger than typical SWE-bench tasks, with an average of 4.0 files and 262 lines changed per fix, and the monorepo contains 63,000 nontest source files and 17 million lines of nontest code. Tasks that change both Java and C++ code had a 55 percent pass rate, compared with 80 percent for Java-only tasks, and Snowflake says the difficulty of crossing the language boundary is a primary factor in agent failure. m:…4522

Snowflake says Cortex Code in REM mode solved 62 of 65 tasks in SnowPR, at 6.34 dollars per trial compared with 8.19 dollars for code mode, while noting that REM mode is currently internal only and not available to customers. Snowflake plans to add rubric-based language-model judges to evaluate design and explanation quality in future versions. In a separate preprint, CRM-Agent-Bench evaluates language agents on multi-turn business tasks in CRM environments using contract-based success criteria, assessing outcome, trajectory, viability, and compliance across 555 stateful, multi-turn tasks spanning nine business domains and 36 capability tags. Strict success rates for the 16 evaluated systems ranged from 64.32 percent to 83.60 percent, success is defined as the strict conjunction of all task-declared checks, and leading systems diverged sharply in cost and per-domain reliability. m:…4522 m:…6562

What measured work changes

The common thread across these four items is measurement on real jobs rather than on generic benchmarks. Vestas measures assignment time against a 72-hour baseline, Dataiku measures whether generated work remains inspectable and permission-bound, Snowflake measures pass rates and cost per trial on real pull requests, and CRM-Agent-Bench measures strict success, latency, and cost across business domains. That shift matters because enterprise buyers do not need another demonstration that an agent can do something once; they need to know how often it succeeds, what it costs per attempt, where it fails, and whether a human can still inspect and override it. The items in this edition give partial answers to each of those questions, and the gaps are instructive: Vestas reports targets, Snowflake reports an internal mode, and CRM-Agent-Bench reports a wide success range across systems. m:…3784 m:…0125 m:…4522 m:…6562