The real bottleneck was not prompting
A lot of the discussion around agents still focuses on prompts. Better prompts, longer prompts, more careful prompts. What made the biggest difference in my workflow was not any of that. It was structure.
I already had a backlog of refactors and ideas in my head, but that is not a real system. It meant too much repetition, too much context rebuilding, and too many tasks started in the wrong order.
My first attempt: a small tool of my own
Before I found Beads, I built a small JavaScript tool that parsed markdown files inside the repo. The goal was simple: make work explicit enough that agents could help without me having to restate everything every session.
I started laying out tasks directly near the code, with priorities like P1, P2, and P3, and with blocking tasks that had to be completed before other work could move forward. That alone changed a lot. Refactoring became more organized. Backlog stopped feeling like one giant blur. Agents had clearer instructions because requirements were written down, prioritized, and tied to actual project movement.
The main improvement was resumability. I could come back later and say, "let's pick up where we left off," and there was already enough structure in place for the session to continue without rebuilding everything from memory.
Where my homemade system hit its limits
That first version worked, but it had real constraints.
Because it lived inside the same repository as the code, it was awkward to use in parallel workflows. It also consumed too much context. Agents did not have a proper built-in way to discover how the system worked, so part of every session still had to be spent teaching the tool back to them.
That was the key lesson: autonomy is not just about having tasks. It is about having tasks in a form agents can reliably discover, understand, and operate.
Then I found Beads
When I came across Beads, I immediately recognized the pattern. It was close to what I had already been converging on, but much more complete.
Beads made the workflow operational. Instead of a homemade layer over markdown, it gave agents a proper system for ready work, claiming tasks, tracking blockers, taking notes, updating statuses, and resuming across sessions. Features like bd prime stood out because they solve a very real agent problem: the tool can teach the workflow back to the agent in a form it can use immediately.
It also includes things I would not have designed myself at first, but that make obvious sense once you work with agents seriously: worktrees, notes, status tracking, and a stronger model for task graphs and session continuity.
What changed after that
Once I adopted this style of working more fully, agents became much more autonomous and much more organized.
I no longer had to repeat myself constantly. My requirements were documented. Priorities were visible. Dependencies were explicit. Work could move forward in the right order. The backlog became something agents could actually help clear instead of just something I kept describing over and over.
That is why this has become a must in all my projects.
Final view
The important shift for me was realizing that agent productivity does not come mainly from clever prompting. It comes from durable, structured work.
Prompts can guide execution. Structure is what makes execution resumable, parallelizable, and trustworthy.
That is the part I had started discovering with a small JavaScript tool and markdown parsing. Beads confirmed it and made the workflow far more efficient.
If you want agents to be more useful, give them better work structure before you give them better prompts.