Durable, Resumable Step-by-Step Workflows with Parallel Agent Steps

[StErMi]

RFC

Summary

Add first-class support for durable, resumable, step-by-step workflows in Flue, where each workflow step can be either programmatic TypeScript logic or agentic work, and completed steps are checkpointed so interrupted or human-blocked runs can resume from the appropriate point.

Background & Motivation

Flue already supports agents, workflows, durable execution, and subagents, but workflows today appear to be finite function invocations rather than resumable step orchestrations. The current model is useful for simple background jobs, but it is harder to model longer-running business processes that need explicit sequencing, fan-out/fan-in parallelism, human review, and selective retry.

Frameworks like LangGraph and Mastra show the value of workflows as durable execution graphs: each step has a clear boundary, persisted state, inspectable output, and resume behavior. Flue would benefit from a similar primitive that still fits its existing TypeScript-first, agent-harness model.

A common user problem is: “I want to run a multi-step AI workflow where some steps are deterministic code, some steps call agents, some steps fan out to multiple specialist subagents, and the workflow can pause for human input or recover from failure without starting from the beginning.”

For example, a review workflow might:

1. Load issue or pull request context with programmatic code.
2. Run multiple specialist agents in parallel.
3. Wait for all specialist results.
4. Stop if any result fails or requires human input.
5. Let a user provide feedback or retry only the blocked step.
6. Continue to the next workflow step once the blocking condition is resolved.

Goals

• Provide durable workflow runs whose progress is persisted at explicit step boundaries.
• Allow workflow steps to run either programmatic TypeScript logic or agentic work.
• Support nested workflows or sub-workflows as reusable steps.
• Support fan-out/fan-in execution where multiple subagents or child steps run in parallel and the parent workflow waits for all results before continuing.
• Preserve successful step outputs so retries do not repeat already-completed work.
• Allow failed, blocked, or human-in-the-loop steps to pause the workflow instead of failing the whole run permanently.
• Allow users or application code to resume a paused workflow with additional input, approval, or custom retry instructions.
• Make workflow state, step status, step outputs, errors, and human-blocking reasons inspectable through the existing run/event APIs.
• Keep the API compatible with Flue’s existing model of agents, subagents, skills, tools, durable streams, and deploy targets.

[ofekron]

The step boundary should persist definition identity, attempt history, and accepted output separately.

A few semantics that prevent painful recovery ambiguity:

• A retry creates a new immutable attempt; it does not overwrite the failed attempt.
• “Completed” means the workflow accepted declared evidence/output, not merely that an agent process exited successfully.
• Human input or approval is a durable event bound to the exact paused step/attempt and expected payload schema.
• Fan-out records the child set before dispatch; fan-in defines whether it requires all, quorum, or an explicit partial result.
• Cancellation distinguishes requested, acknowledged, and outcome-unknown. A late child result cannot silently satisfy a canceled attempt.
• Programmatic side effects require idempotency keys or a reconciliation hook before retry.

Checkpoint the workflow definition digest, effective agent/tool capabilities, source revision, input artifact IDs, and output artifact digests. Resume should fail closed if the executable workflow definition changed incompatibly; offer an explicit migration/new-run path instead of guessing.

I’d also route live execution and restart replay through the same state-transition function. A useful invariant is that the same recorded events yield the same durable workflow state whether processed live, with no UI connected, or after a restart.

I maintain Better Agent (https://github.com/ofekron/better-agent), a local workspace supervising persistent Claude, Codex, and Gemini sessions with delegated work, approvals, and restart recovery. Those operational boundaries are the basis for this feedback. It is source-available and free for non-commercial use; commercial use requires separate permission.

AI-assistance disclosure: this comment was drafted by Codex under the maintainer’s authorization and reviewed in Better Agent.