ACID-Agent

A data-analysis agent where every unit of work is a transaction: explore → decide → generate+execute → validate → commit-or-rollback. Only validated work reaches the workspace; failed attempts are rolled back and leave zero trace.

Implementation of "Agentic Transactions: Towards ACID-Compliant Agent Systems" (Sun, Wang, Li).

What it does

Task ──► supervisor (≤20 units)
           │ per unit:
           │   explore (read-only, bounded)
           │   extract decisions
           │   write & run the unit script
           │   VALIDATION GATE: execution + LLM reflection
           │     + confidence-divergence probes (Qwen3-0.6B, local)
           │   pass  → git commit
           │   fail  → rollback → retry with feedback (≤2)
           ▼
      final answer + full provenance (JSONL trace + Postgres)

The gate rejects code that contradicts the evidence gathered during exploration (token-level log-probability comparison, with vs without evidence), not just code that crashes.

Setup

python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[confidence]"       # core + local scorer
docker compose up -d                 # Postgres + pgvector
python scripts/setup_confidence.py   # download Qwen3-0.6B
opencode auth login                  # backbone (Qwen family)

Run

# one task
python scripts/acid_cli.py run-task "<question>" --agent acid --data-dir path/to/csvs

# A/B evaluation
python scripts/acid_cli.py eval --agent acid --domain biomedical --runs 3
python scripts/acid_cli.py eval --agent baseline     --domain biomedical --runs 3

acid = full transaction machinery. baseline = same model, one raw session, no machinery. Scores and consistency (sqrt(mean per-task variance)) land in results/*.json.

Layout

• acid_agent/graphs/ — supervisor + transaction unit (LangGraph)
• acid_agent/validation.py — the gate
• acid_agent/confidence.py — log-probability probes
• acid_agent/workspace.py — git-backed atomicity (rollback)
• acid_agent/eval/ — KramaBench tasks + graders
• db/schema.sql — runs / units / validations / events