Flowgentic 0.3.0

Flowgentic 0.3.0 is a consolidation release. It preserves the July 2026 work as a documented, tested snapshot: fixes and evidence, no new features.

The LangGraph execution adapter (LangraphIntegration and its asyncflow decorators) is frozen in maintenance mode and gets no new features. The decision follows the substitution experiment, which ran the agentic demo with and without the adapter: its only distinct contribution beyond AsyncFlow and LangGraph was an implicit retry and timeout on every call and graph node, which re-runs scientific work. New applications can bind agents with bind_agent, or run LangGraph on AsyncFlow directly.

Supported configuration

Agent framework	Workflow engine	Backends	Python
LangGraph 1.x	RADICAL AsyncFlow 0.6.0	Local	3.10, 3.11, 3.12

Tested environment, from uv.lock at this tag: LangGraph 1.2.16, langchain-core 1.6.13, langchain 1.4.7, radical-asyncflow 0.6.0 and academy-py 0.3.0. CI runs the test suite on Python 3.10, 3.11 and 3.12.

Install

The wheel attached to this release:

uv venv --python 3.10 flowgentic-env
uv pip install --python flowgentic-env/bin/python \
  https://github.com/stride-research/flowgentic/releases/download/v0.3.0/flowgentic-0.3.0-py3-none-any.whl

Flowgentic is not on PyPI. To run the demos, clone the repository at v0.3.0 and run uv sync --locked.

Links

• Documentation: https://stride-research.github.io/flowgentic/
• Substitution report: docs/evaluation/substitution.md, also on the site
• Changelog: CHANGELOG.md
• The July 2026 demos as they stood before the merge: tag snapshot/radical-demo-2026-08 (demo/radical at dd27bd8b7e6e)
• Archived branches: the 15 branches inactive at this release are kept as annotated tags and deleted from GitHub. Project status explains how to restore one.
  • archive/coordination_overhead_plot
  • archive/example/chemcrow
  • archive/experiment/exp3-benchmarking-hotfix
  • archive/experiment1a
  • archive/feat/abstractions_redesign
  • archive/feat/benchmarking
  • archive/feat/cluster_docs
  • archive/feature/distributed-memory
  • archive/feature/mcp
  • archive/feature/observe-integration
  • archive/feature/telemetry
  • archive/hotfix/benchmarking-engine-and-backends
  • archive/hotfix/benchmarking_redesign
  • archive/refactor/decoupling_tests
  • archive/refactor/worload_old_interface

Known limitations

• DEBUG logging is on by default, and config.yml is looked up in the working directory at import.
• ADR mode is not supported. The demos' --controller adr needs an ADR source checkout (ADR_SRC), because ADR is not released.
• Not validated: using an engine that the caller creates (the integration always creates its own), and any remote or HPC backend. Cancellation reaching in-flight AsyncFlow tasks is validated on local backends only.
• The adapter's default retry policy (3 attempts of at most 30 s each) also wraps graph nodes: a node that runs longer than 30 s fails, and a node that fails with a retryable error re-runs its tasks.
• Token accounting: the agentic demo records no tokens, even in live mode (substitution report, row 5).
• The design-pattern examples need a model provider and credentials. make tests-integration-llm (FLOWGENTIC_RUN_LLM_INTEGRATION=1) runs two of them; CI never does.
• Memory: the trim_middle and importance_based strategies can keep the wrong messages, and summarize can keep one message more than the limit. See Memory.
• The Academy integration (flowgentic.academy) has no tests, and its two examples in examples/academy-integration/ fail at import.

Full changelog: 0.2.0...v0.3.0