ABM (Agent Behavior Mining)

An agentic OS that only shows task boards cannot answer why a path escalated. ABM’s job in this research stack is the observe → event-log bridge for agent runs: turn conversational/tool traces into a process-mining-ready log so evaluation features can mine variants, exceptions, and handoff patterns with standard tools (esp. pm4py).

That serves Evaluations / T38-style work: comparable traces, attributable case IDs, and a path from “agent did something” to “process mining can measure it” — without promoting the demo coffee shop into the system of record.

1. Run the Jupyter coffee-shop notebooks; agents act via LangGraph/LangChain tools and handoffs.
2. Trace collection records activities during interaction.
3. TraceProcessor.process_all_traces(...) converts traces → CSV / JSON / XES under generated_event_log/.
4. Import XES/CSV into pm4py, ProM, or another miner for discovery / conformance / variant analysis.

Related: agent-evaluation · deepeval · hal-the-holistic-agent-leaderboard · topics/32-eval-harnesses

• Use when: you need an XES-oriented export path from multi-agent runs into process mining, or a concrete reference implementation of ABM for eval / governance experiments.
• Skip when: you only need a process-mining library (use pm4py alone), a general agent framework (LangGraph/Crew/…), or production case/ledger semantics.
• Prefer instead: PM4Py for algorithms on logs you already have; keep ABM as the export shape reference.

• Demo domain: coffee-shop O2C is a teaching scenario — patterns transfer, business objects do not.
• Stack lock-in: depends on LangGraph/LangChain + an LLM provider key; Python ≥ 3.13.
• Not the ledger: exported logs are evidence for analysis, not declared company configuration or authority.
• Low star / research repo: treat as paper companion; pin commit if you smoke-test.
• Science shelf: some science/sources quotes still paraphrase landscape/pm4py blurbs — deepen those IDs when Evidence claims need stronger primary text.