What Would an AI-First Project Workflow Look Like?

Beginners
1

As AI agents become better at planning and executing multi-step tasks, I’m wondering if our project management workflows need to change too.

Instead of treating agents as simple assistants, I’d like to see a workflow where:

  • Humans define goals and priorities
  • Agents can own specific tasks
  • Multiple agents can work in parallel
  • Agents share project context
  • Humans review and approve the results

That’s what caught my attention about Sharkly. It treats agents more like participants in a project workflow rather than just another chatbot or automation.

For people building agentic systems with Hugging Face, how are you currently managing task ownership, context, and human review?

Do you see agents eventually becoming actual teammates in development workflows?

4

For my projects, I usually build my own tools and workflows rather than relying on a specific platform. I prefer having direct control over task management, shared context, agent coordination, and human review.

5

What ended up working for us was keeping task state in versioned Markdown files in the repo instead of agent memory. Each file has the task id, owner, status, and module, and every file-changing turn has to anchor to one task.

The human operator is the only one who dispatches work. On parallel runs, a server-side mutex allows one in-flight owner per task, and agents append activity to a shared log without sharing live memory.

6

Hey! I wanted to share some real-world data from my local setup. I’ve been using Qwen 3.8 27B (Q5_K_M) as my main coding model for a couple of weeks to build a complete application, and it’s handling the work surprisingly well.

My agent system runs locally, and I pulled the actual history/database stats from my terminal earlier:

  • Total Sessions: 92 distinct sessions

  • Total Messages logged: 10,627 message turns

  • Raw Text Volume: ~14.37 million characters (~3.5 million tokens) stored in the history database

  • Average Context per Session: ~156,000 characters (~35k–40k tokens)

Because coding workflows hit long context limits fast, I clear/reset the context 4 to 5 times a day to keep the reasoning sharp and

Screenshot at 2026-09-09 17-21-49
Screenshot at 2026-09-09 17-21-492378×1288 345 KB

avoid attention drift. Overall, running Qwen 3.8 27B at Q5 locally has been solid for multi-file app development and heavy tool-use loops.