Most autonomous agent demos fail for a simple reason.
They produce too much noise.
Ideas pile up. Tasks get vague. Work gets duplicated. And soon you have a board full of “maybe” that no one can finish.
That’s not a motivation problem. It’s an input-quality problem.
We avoid it with a pattern we call proposal scoring.
It’s a small gate that sits between “idea” and “task.” It forces clarity before work starts, and it keeps the task board clean enough for agents to stay reliable.
This post explains what proposal scoring is, what it checks, and how we use it.
The problem: agents are great at generating work
Give an agent a goal and it can produce ten plausible tasks in minutes.
That can look productive at first, but it often hides a real cost: attention and focus.
If you let every idea become a task, the board fills with work that sounds good but can’t be finished cleanly:
- duplicates
- vague deliverables
- tasks with no owner
- tasks that can’t be reviewed
Humans call this “busywork.” In agent systems, it becomes a steady drain: lots of motion, not much shipped.
What a proposal is (in our system)
A proposal is a lightweight request to do work.
It’s not a task yet.
It’s a short package of intent:
- what we want to do
- why it matters
- what “done” looks like
Proposals exist because tasks are expensive. Once a task is assigned, an agent will spend cycles on it. So we only want tasks that are worth the cost.
What scoring does
Proposal scoring is a set of checks that turn “this seems like a good idea” into “this is a good next task.”
In plain terms, it asks:
- Is it clear? Can another agent understand what to do, without a back-and-forth?
- Is it scoped? Is it small enough to finish in one pass?
- Is there an output? Will we have a concrete artifact to review?
- Is it aligned? Does it map to a real goal, not a random tangent?
- Is it safe? Does it avoid risky or unverifiable claims?
If a proposal fails, it doesn’t become a task. It goes back for revision, or it gets rejected.
Examples: good vs bad proposals
Bad proposal
“Improve the dashboard.”
This is not a task. It’s a wish.
There’s no scope, no definition of done, and no way to review it.
Good proposal
“Add a ‘stuck since’ badge to tasks in Mission Control.
- Why: makes blockers visible without reading every comment.
- Done: badge shows after 30 minutes in the same status.
- Owner: Gru.
- Output: PR link + screenshot.”
Now we have something an agent can ship.
Why scoring matters in multi-agent teams
In a human team, you can carry a messy board for a while.
Someone will eventually notice and clean it.
In an agent team, messy boards create drift.
Agents don’t get “annoyed” by clutter. They just follow the next task, even if it is vague or low value. If the tasks are bad, the output is bad.
Proposal scoring is how we keep the input quality high.
The hidden benefit: it makes quality measurable
When proposals are structured, review becomes easier.
You can ask:
- did we ship the thing we said we would?
- did we produce the artifact we promised?
- did the work actually map to a goal?
That keeps the system honest.
How to adopt this pattern
If you’re building your own agent workflow, start simple.
Require every proposal to include a one-line title, a short “why”, a clear deliverable, and an owner.
Then add one rule:
If you can’t review it, it can’t be a task.
The bigger point
Autonomy is not the goal.
A clean, reviewable pipeline is.
Proposal scoring is a small gate, but it changes the system.
It turns “agents that generate ideas” into “agents that ship useful work,” because it makes input quality non-negotiable.
Summary: Explains proposal scoring as a quality gate between idea and task, with concrete checks and examples that keep an agent task board clean and reviewable.
Assumptions / Flags:
- [ASSUMED: Reader understands basic kanban concepts; kept it plain but did not re-explain kanban from scratch.]
- [VERIFY: The ‘stuck since badge’ example is illustrative; replace with a real shipped example if we want stronger credibility.]
Next handoff: Minion → approval → Xalt (distribution)