We stopped drawing graphs: an event‑driven runtime for agents

Описание и применение

Reactifact заменяет графы состояния‑ориентированным выполнением, позволяя агентам автоматически реагировать на появление артефактов. Все артефакты версиируются и связываются, обеспечивая воспроизводимость и прозрачную аудиторию.

Задача

Agent frameworks force users to pre‑define a graph of tasks, which is inflexible for knowledge questions that depend on dynamic data and lack a fixed execution path, and they provide little auditability of how answers are derived.

Результат

Reactifact enables dynamic, state‑driven agent execution without explicit graphs, providing deterministic, auditable answers with versioned artifacts and provenance links. A single‑command demo (examples/fintech_audit) shows reproducible variance calculations and audit reports.

Рабочий процесс

• 1. Define artifact types (Question, Evidence, Claim, Calculation, Answer).
• 2. Annotate agents with @produce and Consume(Type) to declare inputs and outputs.
• 3. Create input artifacts (e.g., Question, CSV data).
• 4. Reactifact runtime detects new artifacts and triggers relevant agents automatically.
• 5. Agents perform deterministic Python computations or LLM reasoning, producing new artifacts.
• 6. Each produced artifact records provenance edges to its inputs.
• 7. Run reactifact replay to verify determinism and reproduce results.
• 8. Use reactifact.audit to generate a report with SHA‑256 hashes and provenance.
• 9. Iterate by creating new artifacts or modifying existing ones; only dependent artifacts recompute.

Доказательства

Reactifact is a Celery‑shaped model, not its distributed runtime.

Источник

The audit is a report. reactifact.audit renders the answer with a sha256 per artifact, its producing author, and the provenance edges.

Источник

Run it twice: the hash is the same. Change a fact, and only what actually consumed it recomputes.

Источник