What’s new in Pipewright.

AI build assistance: build and change use cases with the coding agent you already use. The agent writes, Pipewright verifies, a person decides, and no row of your data reaches the agent. Tested end to end with Claude Code on live Google Cloud environments.

Added

• An identity for agents. Each team can have an agent service account with metadata-only roles, which its engineers act as: schemas, row counts and run state, never table data.
• Context packs. pipewright context build writes what an agent reads first: your environment, its rules and its feature catalogue, plus playbooks for common tasks. context check keeps it current in CI.
• A tool server for agents. pipewright mcp gives agents read-and-verify tools; asking for rows is always refused, and every call is logged.
• Judgement checks. validate warns about a random split on time-ordered rows, a date as a feature, features named after the label, and raw features where a feature catalogue exists.
• Gate minimums. A platform can set the least each gate may ask with policies.evaluation.min_gates.
• The review bundle. pipewright review writes a pull request’s description: what a release would change, the settings that moved, and the author’s judgement calls.

Enterprise 1.3.0 adds a hosted tool server per team, aggregates under the platform’s rules, every call audited, EV-16 in evidence packs and an AI agents page in Fleet View. Read the guide.

Found the hard way: 1.2.0 is what building nineteen real use cases end to end on a live environment taught us. Thirty changes, found and tested on live Google Cloud environments, on Agent Platform and on private GKE.

Added

• Serving gated in CI, or automatic. approvals.deploy_model: ci deploys a new champion once the GitHub environment’s reviewers approve.
• How to call a deployed model. deploy-model ends with the endpoint, the URL, the permission needed and a sample request.
• GPU limits and GPU serving. The platform sets the types and counts; a use case serves on a GPU within them.
• Scale to zero for endpoints that are rarely called.
• The decision trail. pipewright lineage shows who approved, promoted and deployed each version.
• Exact replays. A rows fingerprint on every snapshot, and splits that are the same every time.
• Promotion in CI. ci init writes a promotion workflow for environments fed by promotion.
• A failed run is an event you can be emailed about, with the failed steps and their errors.
• Re-check a version against today’s gates, with the champion’s slices and fairness on the same rows.
• Forecasts gated per series, a base image per use case, tests before a release, image cleanup, and --json for scripts.

Fixed

• Online serving fills an empty feature as training did, and a promoted model reads its own environment’s online store.
• Performance monitoring joins each outcome to the score made for its cut-off.
• A refused retrain records the gates’ reasons and re-opens its case.
• A release after an upgrade carries the new configuration as a new release ID.

More models on the governed path, and less manual work keeping them healthy.

• BigQuery ML, governed. Classification, regression, k-means segments, PCA, ARIMA_PLUS forecasts and imported TensorFlow and ONNX models, with byte limits.
• The closed retrain loop. Drift in production opens a case; the candidate trains in test and returns only with a person’s approval.
• Failed retrains, handled. Recorded on the case, retried on their own when transient, by a named person otherwise.
• Online features at serving time, from the same definitions training uses.
• Approvals in GitHub and fairness gates that stop a model treating groups unequally.
• Lineage: pipewright lineage in the core; drawn in Fleet View in Enterprise.

The first stable release: use cases in YAML, write-once releases, the shared launcher, gates and champions, monitoring, retrain decisions with approver quorums, promotion by digest, Ray and features at a point in time, synthetic data, and private GKE. From 1.0 on, every change is additive.