Zenflow streamlines Al-first engineering with specification-driven workflows, parallel agents, and built-in verification - so you can ship production-grade software.
Andrew here. While building our IDE extensions and cloud agents, we kept running into the same problem many of you probably face when using coding agents in complex repositories: agents getting stuck in loops, over-apologizing, and burning time without making real progress.
We tried to paper over this with scripts, but juggling terminals and copy-paste prompting quickly became painful. So we built Zenflow - a free desktop tool for orchestrating AI coding workflows.
It handles the things we kept missing in standard chat interfaces:
Dynamic Workflows: Workflows are defined in simple .md files, and agents can dynamically rewire the next steps based on what they discover mid-execution.
Spec Driven Development: Use formal specs to guide agents, ensuring the implementation matches your architectural intent before a single line of code is written.
Cross-Model Verification: Have Codex review Claude’s output, or run multiple models in parallel to see which one handles a specific codebase or task best.
Blast Mode (Multi-Model Inference): Run up to four different models (Claude, GPT, Gemini, Codex) on the same task simultaneously. Compare their outputs side-by-side and pick the best result.
Parallel Execution: Run multiple approaches on the same backlog item simultaneously mixing human-in-the-loop workflows for hard problems with faster “YOLO” runs for simpler tasks.
Project-Level Kanban: Track and manage all agent work through project lists and kanban-style views, not scattered terminal windows.
What we learned building Zenflow
After running 100+ experiments on SWE-Bench and private datasets, we found that models are increasingly overfit to public benchmarks. Real-world success doesn't come from "smarter" models alone; it comes from the "Goldilocks" Workflow just enough structure to prevent loops without over-orchestrating the creativity out of the AI.
We’ve been dogfooding this heavily to build our own IDE extensions, and we’d love to hear how it handles your toughest repos.
Zenflow is free to use and currently supports Claude Code, Codex, Gemini, and Zencoder.
Interesting. We’ll review it with the team — we’re actually building several AI applications.
A very interesting and ahead of its time product for working with AI agents at the task, agent and workflow level. Works quite well right out of the box with its own agent, not to mention all the features. The product is still a bit lacking in flexibility/configurability, so I wish the team to gather all the useful feedback and make the product even better!
I'm excited to see how people incorporate this into their real workflows.
Zenflow sounds like a powerful way to professionalize AI coding! I’m really curious if the "built-in verification" can actually catch subtle logic bugs that AI often misses. Since it uses parallel agents, I wonder how it keeps them all synced without causing conflicts in the code. Is the specification-driven workflow flexible enough to handle sudden changes in a project’s direction? I’d also love to know if it integrates easily with the tools we already use or if we’d need to move our whole process over to their platform!
Well done on this release.
This looks very interesting. I've always believed that we should be running agents as we would run a development, and that means working on a specification to work against first of all - I also believe agents should be collaborating with you on that specification and critiquing it as well.
Does your system also do that as well (run through the full SDLC process) or does it start with a fully defined specification?
Big fan of you @andrewsthoughts and your whole Zencoder team! Congrats on your latest launch!
I've been using it for a couple of weeks now and am now just babysitting the AI as it builds the product! Code quality is very good!
If a team is already using Cursor/Claude Code/Codex directly, what’s the clearest reason to add Zenflow rather than just writing internal scripts or stricter prompting guidelines—and what’s the switching cost you had to eliminate to make that decision easy?
Huge congrats! Zenflow’s multi-agent orchestration and built-in verification feels like a game-changer for scalable AI engineering. I'm excited to try real workflows beyond vibe coding. Good luck!
Exciting launch, crew!
You guys are really taking spec-driven development to the next level. I'm excited to see how it pans out. Let's gooo!
Multi-agent “blast mode” and dynamic rewiring is powerful, but at scale the pain is non-determinism: agents race, loop, and produce conflicting diffs that are hard to replay or audit.
Best practice is a reproducible execution harness: sandboxed per-agent workspaces/branches, deterministic step graph with idempotent tools, and mandatory verify gates (lint + minimal tests) before merge, with full traces for replay.
How does Zenflow represent and version the workflow state, and can it enforce conflict-free patch application plus automatic rollback when verification fails?
🚀 Hey Product Hunt!
Andrew here. While building our IDE extensions and cloud agents, we kept running into the same problem many of you probably face when using coding agents in complex repositories: agents getting stuck in loops, over-apologizing, and burning time without making real progress.
We tried to paper over this with scripts, but juggling terminals and copy-paste prompting quickly became painful. So we built Zenflow - a free desktop tool for orchestrating AI coding workflows.
It handles the things we kept missing in standard chat interfaces:
Dynamic Workflows: Workflows are defined in simple .md files, and agents can dynamically rewire the next steps based on what they discover mid-execution.
Spec Driven Development: Use formal specs to guide agents, ensuring the implementation matches your architectural intent before a single line of code is written.
Cross-Model Verification: Have Codex review Claude’s output, or run multiple models in parallel to see which one handles a specific codebase or task best.
Blast Mode (Multi-Model Inference): Run up to four different models (Claude, GPT, Gemini, Codex) on the same task simultaneously. Compare their outputs side-by-side and pick the best result.
Parallel Execution: Run multiple approaches on the same backlog item simultaneously mixing human-in-the-loop workflows for hard problems with faster “YOLO” runs for simpler tasks.
Project-Level Kanban: Track and manage all agent work through project lists and kanban-style views, not scattered terminal windows.
What we learned building Zenflow
After running 100+ experiments on SWE-Bench and private datasets, we found that models are increasingly overfit to public benchmarks. Real-world success doesn't come from "smarter" models alone; it comes from the "Goldilocks" Workflow just enough structure to prevent loops without over-orchestrating the creativity out of the AI.
We’ve been dogfooding this heavily to build our own IDE extensions, and we’d love to hear how it handles your toughest repos.
Zenflow is free to use and currently supports Claude Code, Codex, Gemini, and Zencoder.