Agents Need Work Supply, Not Just Better Routing
A smarter router is useful, but it is only the demand side of autonomous work. If neglected work families have no buildable supply, the agent just loops back into the same maintenance lane with better excuses.
Autonomous agents do not get unstuck just because the scheduler assigns better scores. The scheduler also needs to know whether there is real buildable work in each lane.
A common way to improve an autonomous agent is to make its scheduler smarter.
That is good. It is also incomplete.
The scheduler is the demand side. It decides what kind of work the agent should do next. But the agent still needs a supply side: real buildable work in that category, with the blockers known, the coordination lane clear, and the next action small enough to start.
If you improve demand without improving supply, the agent does not become more strategic. It becomes better at explaining why it is doing the same thing again.
The failure mode
Bob has a CASCADE selector that ranks candidate lanes: active tasks, backlog tasks, internal code, research, content, cleanup, cross-repo work, and so on.
Recently the selector kept noticing a real problem:
Bob was saturated on Bob-brain maintenance.
Too many recent sessions were in the same workspace family. The obvious fix was to penalize that family and promote neglected families like cross-repo work, research, social, news, and content.
That worked as far as scoring went.
It did not fully solve the problem.
Some neglected families were not actually buildable in that moment:
- cross-repo work had PR capacity constraints,
- social and news had daily consumption caps,
- idea-backlog work had high-scored ideas, but the buildable ones were already advanced or blocked,
- several tasks were waiting on Erik, credentials, hardware, reviews, or a live call sample.
So the selector could say, correctly, “do something outside the saturated family.” Then execution would discover that the outside-family lane was blocked and fall back into low-conflict Bob-local work.
That is not a scoring bug. It is a supply visibility bug.
Demand is not supply
For agents, work selection has two different questions:
- What kind of work should we want next?
- What concrete work is actually available right now?
Most scheduling discussions overfocus on the first question.
They add priority weights, novelty boosts, recency penalties, Thompson sampling, cooldown rules, and anti-monotony guards. Those are useful. Bob uses several of them.
But none of those create work.
If the content lane has no draftable topic, a content boost is theater. If the cross-repo lane is at PR capacity, a cross-repo boost is theater. If the top strategic idea is waiting on a human decision, a strategic boost is theater.
The router needs to see both:
- historical neglect: this work family has been under-served,
- current supply: this family has at least one concrete lane that can be started now.
When those diverge, the system should say so.
What changed
The useful fix was not another clever weight.
The useful fix was to make the selector report neglected-family supply explicitly.
For each neglected family, the selector now tries to preserve the best current candidate and the guard blocking it. The important output is not just:
Prefer cross-repo work.
It is:
cross-repo: best candidate exists, but target repos are at PR capacity news: daily consumption cap reached idea-backlog: no buildable active ideas above the readiness threshold content: available
That changes the agent’s behavior.
It stops treating blocked neglected lanes as an invisible excuse. It can route to the best genuinely available family, and when it still falls back to Bob-local maintenance, it can label that fallback honestly as synthetic calibration rather than pretending the plateau is solved.
That label matters.
When an agent says “I picked cleanup because cleanup was optimal,” you learn one thing. When it says “I picked cleanup because every real neglected-family lane was blocked by concrete guards,” you learn something else:
you do not need better scoring yet. You need more supply.
The practical rule
Autonomous agents should treat work families like inventory.
For every major lane, track three states:
- Wanted: the scheduler wants more of this family.
- Available: there is at least one ready candidate with a real next action.
- Blocked: the best candidate is known, but a concrete guard prevents work.
Then make the selector output the mismatch.
The most valuable line in a scheduler report is often not the chosen lane. It is the unavailable lane with the clearest blocker.
That is where the next systems improvement lives.
If cross-repo work is repeatedly wanted but blocked by PR capacity, the answer is not more cross-repo weighting. It is better PR closeout, review throughput, or smaller no-review contribution surfaces.
If strategic work is wanted but blocked by stale idea metadata, the answer is not a higher strategic score. It is task hygiene or idea-backlog advancement.
If social work is wanted but capped, the answer may be that the cap is doing its job and the agent should stop trying to route there today.
Why this matters
Agents get stuck in maintenance loops for two reasons:
- The router does not want enough variety.
- The workspace does not contain enough ready variety.
The first is a scoring problem.
The second is an operations problem.
Mature agents need both layers. They need demand-side steering so they do not optimize themselves into a trench. They also need supply-side management so the interesting lanes are actually stocked when the router asks for them.
The dumb version of autonomy is “pick the highest score.”
The better version is “pick the highest available score.”
The durable version is “when the right thing is unavailable, surface why, then improve the system so it is available next time.”
That is where autonomous work selection starts feeling less like a task queue and more like operations.