Skip to main content
Pollack AI Lab
Build · Evaluate · Improve Agentic Software
The aim is simple: build agents, measure what they do, and understand what makes them better. The lab contains the methods, software, and experiments behind that work. By Mark Pollack, creator of Spring AI.

Agento Forge

Methodology and executable toolkit for building and evolving agentic software. Six phases, with evaluation and retained learnings as first-class artifacts.

AgentWorks

Reusable infrastructure for running, observing, evaluating, and experimenting with agents on the JVM.

Experiments

Controlled experiments on what changes agent behavior and outcomes, with the variants and data published alongside each result.

​
Highlights

What’s genuinely new and cool across the lab right now.

Run a workflow step in Anthropic's cloud

ManagedAgentStep hands one step of your workflow to Claude Managed Agents — a full hosted sandbox with bash, files, and web. Same graph; swap a single step between local JVM, Temporal, or Anthropic’s cloud.

Speak the latest Agent Client Protocol

Tracks ACP 0.14.0 — elicitation (agents ask you for structured input), session fork & resume, and provider config — dropping a Java agent straight into Zed, JetBrains, and VS Code.

Evaluate any agent, any framework

Agent Judge verifies whether work from Spring AI, LangChain4j, Koog, or any CLI agent satisfies its goal — combining executable checks, LLM and RAG evaluation, and explicit jury policies. Read the deep dive →

Provably CVE-clean dependencies

The AgentWorks BOM ships only behind a 9/9 convergence gate (plain / Boot 3.5 / Boot 4) — every consumer gets one coherent, vulnerability-free dependency set, verified in CI, not by hand.

Trace Claude and Gemini in one format

Agent Journal captures Claude and Gemini CLI runs into one portable trace + cost schema — one analysis layer across runtimes, no per-vendor lock-in. New in 1.5.0.

​
Active release train

The whole suite ships as one coordinated train, pinned by the AgentWorks BOM — currently 1.18.0, spanning a dozen libraries all shipped within the last month.

​
Core Projects

ACP Java SDK

Agent Client Protocol — build and consume agents for Zed, JetBrains, VS Code

Agent Workflow

Multi-step agentic pipelines with typed context, quality gates, and portable runtimes

Agent Judge

Evaluation with deterministic, command, and LLM judges

Agent Client

Claude, Gemini, and Codex as actively verified Spring services

Agent Journal

Behavioral trace capture for analysis and observability

Agent Bench

Benchmarking on real enterprise Java tasks

​
Latest Experiments

Code Coverage v3 — The Exemplar Effect
When existing tests use older patterns, skills can’t override them — the codebase is the agent’s primary teacher. 2 variants on Spring PetClinic.
Apr 2026 · Complete · Details →
Code Coverage v2 — Skills vs Knowledge Bases
Do structured skills beat flat knowledge injection? Skills + pre-analysis cut steps 31% with no quality loss. 7 variants.
Mar 2026 · Complete · Details →
Code Coverage v1 — Knowledge Injection Baseline
9 variants testing progressive knowledge injection on Spring Boot test generation.
Feb 2026 · Complete · Details →

​
Where this work leads

The software and the method get applied to real products. Bud Spring is a family of agents for controlled software work, and a sibling project, Bud DDD, applies the same method to domain-driven-design review without tying it to Spring, Java, or any single framework. Agento University is a longer-term exploration of a visual operating environment for a fleet of agents that each own a codebase. These are examples of where the work leads, not the organizing idea of the lab.