Project status¶

Maintenance mode¶

Flowgentic 0.4.0 is a repair release in maintenance mode. 0.3.0 preserved the July 2026 work as a documented, tested snapshot, and 0.3.1 fixed defects that 0.3.0 documented. 0.4.0 removes features that 0.3 shipped broken or unsafe, or that did not serve Flowgentic's purpose, and adds acceptance of agent outputs and one trace across decisions, tasks and model calls. The changelog lists each change, and what replaces each feature it removes.

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. Without it, the demo kept the same results, task identity, cancellation, service reuse, backend placement, tool selection and handoffs, in about the same code and with 32 fewer third-party packages. The adapter's only distinct contribution was an implicit retry and timeout on every call and graph node, which re-ran scientific work; 0.4.0 removes them.

bind_agent, which binds an agent component to an AsyncFlow task, remains a documented recipe: Augment an AsyncFlow application. accept, new in 0.4.0, accepts or rejects an agent's output on explicit checks, with re-asks and a deterministic fallback: Acceptance. A LangGraph application can also run on AsyncFlow without Flowgentic: LangGraph on AsyncFlow directly. So can an application that calls Academy agents from its tasks: Academy actions as AsyncFlow tasks, which replaces the flowgentic.academy module that 0.4.0 removed. 0.4.0 also removes the memory manager and the mutable graph, which served none of this. LangChain's middleware bounds an agent's history. LangGraph routes within a run, and a graph that should change between runs is built anew as a StateGraph; see the changelog.

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: LangGraph 1.2.16, langchain-core 1.7.0, langchain 1.4.7, radical-asyncflow 0.6.0 and academy-py 1.0.0. CI runs the test suite on Python 3.10, 3.11 and 3.12, and on Python 3.10 with the lowest versions the dependency ranges allow; a weekly workflow runs it, and builds the docs, on the newest releases within the ranges. The package requires langgraph>=1.1.5,<2.0; langchain-core, langchain, langchain-openai and langchain-ollama at >=1.0,<2.0; and radical-asyncflow>=0.6,<0.7, because before 1.0 an AsyncFlow minor release may break; and opentelemetry-api>=1.43,<2 for its spans. Its examples extra adds academy-py>=1.0,<2 for the Academy example, and the OpenTelemetry SDK, OpenTelemetry's GenAI instrumentation for LangChain and OpenAI (1.3b0, below 1.4) and rhapsody-py[telemetry] for the demos' traces.

Known limitations¶

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 recipe for calling Academy agents from AsyncFlow tasks is tested with Academy's local exchange and AsyncFlow's local backend only; see Academy actions as AsyncFlow tasks.
AsyncFlow 0.6.0 has no per-task retry or timeout, so applications wrap the task calls that need them. See Retries and timeouts.
Tracing (see Telemetry):
- A model call inside an AsyncFlow task is a sibling of the task's span, not its child.
- AsyncFlow block spans hang from the session span, not from the caller's.
- AsyncFlow 0.6.0 labels every task span with the default backend.
- No cost that a provider reports is recorded.
- Acceptance's evaluation events are log records, emitted through Python's OpenTelemetry Logs API, which still lives in the module opentelemetry._logs. An application sees them only once it configures a logger provider.
- RHAPSODY's Dragon-vLLM endpoint reports word counts as token counts.
Live runs of the design-pattern examples need a model provider and credentials. CI runs three of them with a scripted chat model in place of the provider (tests/integration/test_design_patterns_offline.py), which shows that they complete, not what a live model makes of them. make tests-integration-llm runs two of them live, given OPEN_ROUTER_API_KEY, and the LLM examples workflow does the same on GitHub when a maintainer starts it. The interactive chatbot runs in neither.

Retired¶

Flowgentic no longer plans the framework × backend matrix that earlier releases announced (CrewAI, AG2 and the OpenAI Agents SDK, on Pegasus, Parsl and Academy), nor universal agent abstractions of its own.

The abstractions' prototype is kept in archive/feat/abstractions_redesign, and the AGENT_TOOL_AS_MCP flow type in archive/feature/mcp; see Archived branches.

Archived branches¶

The branches inactive at the 0.3.0 consolidation are archived as annotated tags, archive/<branch>, on their last commit, and deleted from GitHub. No commit was lost.

Tag	Commit
`archive/hotfix/benchmarking_redesign`	`45268b5a73e2`
`archive/hotfix/benchmarking-engine-and-backends`	`44d5a4132ee7`
`archive/experiment/exp3-benchmarking-hotfix`	`8bf7270425d6`
`archive/refactor/decoupling_tests`	`0940607b41ca`
`archive/feat/cluster_docs`	`feecf47f54b4`
`archive/experiment1a`	`613b536586c8`
`archive/coordination_overhead_plot`	`5c0905438b68`
`archive/feat/abstractions_redesign`	`1ec100eaeb10`
`archive/refactor/worload_old_interface`	`9a1d45b02d87`
`archive/feat/benchmarking`	`4289a01d9e25`
`archive/feature/telemetry`	`7ff024df1ed1`
`archive/feature/distributed-memory`	`40fffaf3f861`
`archive/feature/mcp`	`73333a4a7652`
`archive/feature/observe-integration`	`0b121d045415`
`archive/example/chemcrow`	`1972bcc0c656`

Four more branches, hotfix/deafult-settings, docs/memory, feature/memory and features/supervisor, were already part of main and were deleted without a tag.

To work on an archived branch, fetch the tags and create a local branch from one:

git fetch origin --tags
git switch -c feature/mcp archive/feature/mcp

A maintainer restores the branch on GitHub by pushing the tag's commit to it: git push origin 'archive/<branch>^{commit}:refs/heads/<branch>'.

The tag snapshot/radical-demo-2026-08 marks the July 2026 demos as they stood before they were merged into main: commit dd27bd8b7e6e (19 August 2026), the tip of the former demo/radical branch, which was deleted on 22 October 2026.

Earlier releases¶

Tags v0.1.0 (23 September 2025) and 0.2.0 (2 October 2025) predate the changelog.