From AI Agents to AI Teammates: What Should the Workflow Look Like?

Beginners
1

AI agents are getting better at planning, using tools, and completing multi-step tasks. But I think the next challenge is less about making agents autonomous and more about making them useful teammates.

A practical workflow could be:

Human defines the goal → agent executes tasks → human reviews → agent iterates.

This becomes especially interesting when agents need to share context, track tasks, and hand work back to humans instead of operating in isolation.

I’ve been exploring this idea with Sharkly, which focuses on bringing AI agents into project workflows alongside human teammates.

For those building agentic systems with Hugging Face models, how are you approaching human-in-the-loop workflows today?

Do you treat agents as tools, assistants, or actual participants in your workflow?

4

“Teammate” should mean more than conversational tone. A useful test is whether the agent can own a bounded responsibility, expose uncertainty, return evidence with its work, and preserve the team’s latest decision instead of its own stale interpretation. Irreversible authority still needs an explicit boundary.

5

Hi Eric,

I have been experimenting with software development in a production-like workflow. I wrote about some of my early experiences on medium.com a while ago. What is currently working:

I use Claude to turn my idea into a prompt/context document for the coding agent and a csv with the initial product backlog (epics, user stories, acceptance criteria, technical notes). I upload these into Jira, but Claude can also use Atlassian’s MCP server to do it directly. That will eat up a lot more tokens though.

Next step is to put an agent to work. I use Visual Studio Code with a Copilot subscription. I have experimented with various local models, but for complex, long range tasks I use cloud models (either Claude or GPT in auto mode). The prompt contains the development workflow and instructs the coding agent to pull the next task from the backlog, create a branch, do the work, update Jira, create a PR and wait for review. Depending on the complexity, you can review one user story at a time, or an Epic, or just let the agent keep coding.

This workflow works well, I have a fully dressed backlog, traceability in Jira and Git, and built in review steps (by me and by AI on GitHub in the PR flow). It does take much longer than just writing prompts as you go, and it considerably ups your token usage rate…

I haven’t tried expanding this workflow to multiple agents working concurrently, but this is definitely something to try. I would also love to see the workflow tied in to a CI/CD pipeline, but that is a bit beyond me on my own and with my limited hardware :hugs:

I have had brief looks at Cline and Hermes, but not put the time in to get a good understanding and my initial experience with local agents + kanban have been disappointing to say the least. Since then Qwen 3.8 has been released, that is a much more usable model than what was available before. Jira is industry standard backlog management, but it’s also collapsing under its own weight with legacy and features. I would definitely be interested in a lightweight alternative.