September 2026 Summaries

Chalk introduces SQL and DataFrame interfaces for its federated query engine, designed to help ML and AI teams query, transform, and serve data spread across databases, warehouses, streaming systems, APIs, and external services without manually managing data movement. Chalk SQL offers a declarative approach to combining sources such as Postgres and BigQuery, while the Python-oriented Chalk DataFrame library supports iterative workflows familiar to pandas and Polars users; both compile into the same optimized execution plan. The engine can invoke external functions and services, including tools such as Jev, while pushing compatible filters, limits, and projections down to underlying sources to reduce scanning and data transfer. Underneath, Chalk uses and extends Meta’s Velox C++ execution engine with custom operators and integrations for more than 10 databases, aiming to support both high-throughput analytics and low-latency feature serving. By sharing a unified planning and execution model across SQL, DataFrames, and supported Python feature logic, Chalk seeks to shorten the path from exploratory analysis to production inference and agent-context workloads.

Chalk built MealBot, an internal agent that answers employees’ questions about catered lunches, retrieves menu information from Zerocater, supports Slack interactions, and encourages meal ratings to improve future orders. To limit its scope and prevent data leaks or unauthorized actions, the agent runs in Chalk Sandboxes with outbound network policies allowing access only to OpenAI, Zerocater, and Slack, including restrictions on specific API endpoints. Chalk’s authentication injection keeps real API credentials inaccessible to the agent by adding secrets only to valid approved requests. The company presents MealBot as an example of how production agents can be deployed securely within a customer’s VPC while maintaining control over execution environments, permitted tools, network access, and credentials.

Chalk has introduced Chalk Notebooks, a hosted notebook environment for agent-assisted machine learning analysis that runs within a customer’s cloud environment and connects directly to production data and deployments. The product combines Python, SQL, markdown, a CLI, and Model Context Protocol support, enabling agents to perform federated queries across databases, warehouses, streams, APIs, and feature stores while preserving point-in-time correctness and allowing comparisons between production and isolated Chalk branches. In a loan-default investigation demonstration, an agent identified a group of Florida pawn shops associated with unusually high defaults, created time-valid merchant features, retrained and evaluated models, performed walk-forward validation to avoid hindsight bias, and recommended a merchant rule, a refreshed model, and additional features based on their confidence levels. The company emphasizes that notebooks preserve the reasoning, evidence, and data context needed to reproduce agent-driven ML decisions, while governance controls restrict agent access through scoped credentials, sandboxed compute, row and column policies, controlled egress, and disabled production write access; agents can investigate and test changes, but humans remain responsible for promoting them to production.

Chalk has introduced Chalk MCP Server and Chalk Assistant to integrate AI agents directly into machine learning development workflows, aiming to reduce the manual effort involved in gathering context, investigating models, engineering features, and validating changes. Using the Model Context Protocol, the MCP Server lets agents access Chalk capabilities such as reading feature definitions, running queries, analyzing data with Chalk SQL, diagnosing errors, engineering and backtesting features, and investigating model behavior. Chalk Assistant provides an in-product interface where users can connect an agent through their own model API key, ask it to examine models, features, or queries, and review its reasoning and proposed changes. The company positions these tools as a foundation for “agentic machine learning,” in which agents actively support continuous experimentation, faster iteration, and proactive model improvements within production ML systems.

Agentic machine learning uses agents to automate the iterative model-development cycle of proposing hypotheses, implementing changes, evaluating results against metrics, and retaining or discarding modifications while humans set objectives and approve production changes. It can increase experimentation throughput, investigate more prediction failures, and help teams support additional models without proportional staffing growth, but differs from AutoML by generating or reframing hypotheses rather than only searching predefined spaces. Its central risks are not simply coding errors but misleading evaluation results caused by feature leakage, point-in-time join mistakes, contamination of held-out data, unavailable production features, and latency constraints, all of which can improve offline metrics without producing real-world gains. Trustworthy deployment therefore requires human review calibrated to risk, evaluation of both model outcomes and agent trajectories, reproducible traces, and access to the same source data and feature definitions used in production. The piece argues that a data layer providing point-in-time-correct training data, shared training-serving feature definitions, lineage, governance, and federated access to operational sources is essential, presenting Chalk’s Context Engine as infrastructure intended to support these requirements.

Real-time numerical context can make agent decisions more responsive than static retrieval or precomputed aggregates, according to a Chalk demonstration using poker strategy. Starting with a Game Theory Optimal chart as a baseline, the example compares an opponent’s lifetime behavior, current-session statistics, and short trailing windows, showing how a recent shift toward aggressive play can alter a recommended action from calling most often to raising 71% of the time. Chalk’s windowed feature definitions are presented as a way to calculate metrics directly from raw data at the moment of a request, without separate streaming jobs, rollups, or materialized views, while its deployed agent uses a sandboxed function and a single query tool to retrieve features. The approach is positioned for low-latency, high-throughput use cases where recent behavior matters, such as fraud detection, by replacing poker-specific entities with cardholders, transactions, and model expectations.