This quarter, we focused on how engineers and AI agents work together in Chalk. Agents can investigate problems and run experiments, with engineers reviewing the results and controlling access to models and tools.
Here’s what’s new.
Bring agents into your development workflow
Chalk MCP Server connects agents to Chalk’s data and development tools. Agents can query connected sources, invoke functions, run notebook cells and inspect logs and traces to test ideas against your actual environment.
Chalk Assistant is an AI assistant built directly into the dashboard. Ask it to investigate an unexpected result, and it can run queries and record its findings in a Chalk Notebook. Python, SQL, and Markdown cells keep queries, training experiments, results, and explanations together for engineers to review.
Turn experiments into better models
Chalk Model Training lets teams train models on the features they already use in production. Define the training code, dataset, and compute requirements; Chalk manages CPU or GPU execution, with versioned inputs, saved checkpoints, and tracked metrics. You can now deploy those models directly from the dashboard.
An agent investigating a rise in loan defaults can use a Chalk Notebook to identify the affected segment, develop features, launch training, and compare a candidate with production. The engineer gets both a proposed improvement and the evidence behind it.
At Grid, Chalk Model Training helped cut model development time by more than half and increase training runs from roughly 10 to 30 per week.
Build agents with control over their tools and models
Chalk Model Gateway gives teams one place to control model usage and reliability across providers. Set budgets, rate limits, and fallbacks once instead of rebuilding them in each application. It runs within your existing Chalk deployment, with model-call traces alongside your queries to help investigate behavior.
Chalk MCP Gateway lets agents access approved tools using their Chalk identity, without distributing backend credentials. Policies control which tools agents can call and with which arguments, while an audit trail records each operation.
Use familiar and intuitive interfaces
Chalk SQL and Chalk DataFrame give teams familiar ways to work with Chalk’s federated query engine. SQL, DataFrame operations, and Python feature definitions use the same underlying engine, connecting exploration with the logic used for production features.
A data scientist can join transactions in Postgres with historical records in Snowflake in one Chalk SQL query. An engineer can express an iterative transformation with Chalk DataFrame. An agent can use those same interfaces to explore data and test a hypothesis. Chalk plans the computation across connected sources and pushes eligible work to the source.
Chalk LSP and IDE plugins bring diagnostics, code navigation, and typechecking into your editor. Iceberg REST catalog support adds more ways to connect lakehouse data, while SQL linting in Chalk MCP Server lets agents check queries for planning errors before running them. Together, these additions help engineers and their agents explore data, validate queries, and move from investigation to implementation using the same engine.
Talk to an engineer about bringing AI to production with Chalk.
