Cheap subagents need a result and a stop.

01Cheaper workers still need a useful result

On 5 October, Reddit user Imanflow described moving from long Opus sessions to separate topics, Markdown handoffs and model assignments by difficulty. Their usage screen then attributed 82% of usage to subagent-heavy sessions. They were already choosing cheaper workers. The author explicitly said they had not properly compared the approaches: 82% is a share of usage, not a measured saving or increase against the same work in one session. ⁶

Haiku 5.5’s 7 October announcement gave teams another reason to try bounded delegation. Anthropic positions it for summaries and subagent tasks alongside larger models, with different published rates above 100,000 prompt tokens. That makes each request’s size and assignment part of the comparison. A lower worker rate alone cannot answer Imanflow’s question. ¹

Take a hypothetical checkout bug. A lead agent asks a smaller scout to find the discount calculation, name the relevant paths and cite the code. The scout can return a useful map, a wrong map or an honest failure. Its assignment ends when the lead can judge that evidence or the agreed search limit is reached. It does not own the patch.

02A handoff needs an acceptance check

The lead opens the cited paths and follows the discount calculation. A correct map lets implementation begin. A stale path, an unrelated helper or a claim without code sends the task back for a bounded retry or a different search. The scout has not saved work if the lead has to repeat all of its discovery before trusting a single line.

That acceptance step also defines the comparison. Both the existing route and the proposed scout route must locate the same behavior to the same standard. Counting the number of returned summaries would reward activity without establishing usefulness. Counting accepted maps keeps the result attached to the developer’s actual question.

A second practitioner reached a narrower model-routing decision. In an edit to an 8 October Reddit post, visible when checked on 19 October, emarkosov reported reviewing 129 recent subagent spawns and finding only two Explore runs. Most were implementation tasks they wanted to keep on Opus. The author therefore chose to move research workers individually to Haiku rather than force every worker onto it. Those counts describe one person’s workload, without an independent cost or correctness comparison. ⁷

That distinction matters for our scout: discovery, implementation and review have different accepted outputs. Change the model for one role, then measure whether its handoff still meets the standard. The author’s correction identifies the assignment mix that a blanket switch would hide.

03Denied is a task state, too

Suppose the checkout scout tries to edit a file despite its read-only role. The intended result is that the write does not occur and the lead receives a usable explanation. Claude Code 2.1.295 added onFailure: "block" for command and HTTP hooks: startup failures, timeouts and unexpected exits can block the action. A mandatory checker can therefore have an explicit failure policy. ²

The following release, 2.1.296, fixed managed PreToolUse hooks using continue: false, and managed prompt hooks, refusing a call without ending the turn. It also fixed specified headless-session cases that started folder-disabled MCP servers after a directory change or plugin reload. These are dated vendor fixes; they show why the blocked action and the resulting session state need separate observations. ³

For the scout, a refused write must not turn into a completed discovery result simply because the turn ended. Nor should a checker outage disappear inside an automatic retry. A small evaluation can deliberately attempt a prohibited edit, then make the required checker unavailable. The file diff and the reported task state reveal whether the configured workflow handles both cases as intended.

04Keep the question attached to the task

Now the lead has found two competing discount rules and needs a human choice. Plannotator’s 0.28.8 Inbox preview collects agent questions, files and reviews in one local window, with replies directed to the originating session. It can also show that an attached file changed after it was sent. Those details matter when the developer reviews several tasks at once. ⁴

In our example, the question should identify the checkout task and the revision under discussion. If the lead’s session ends before the answer arrives, the workflow needs a visible outcome: delivery to the intended session, a recoverable pending answer or an explained failure. A test that sends a reply after a restart can investigate this behavior. We have not run that test here.

Remote access adds another condition. Cursor’s 6 October release lets its iOS app steer local agents, but the work still runs on the computer, which must remain on and online. ⁵ Answering from a phone does not move the checkout task to a persistent remote runner. Work awaiting approval should stay pending while independent tasks continue under their own permissions.

05Count the work that reaches acceptance

For a trial, run the same checkout discovery through the existing route and the scout route. Keep every attempt: model requests, accumulated context, retries, the lead’s review and any repeated search. Then compare the cost and elapsed time of an accepted map. A lower token rate is one input to that calculation; a map that needs extensive repair can consume the difference.

Prompt length can change during the task as tool results and conversation history accumulate. The relevant size is the request actually sent on each turn, including its billing category and any exposed cache usage. The short instruction at the beginning cannot establish which pricing tier later turns use.

The trial should answer Imanflow’s unresolved comparison: what did the same accepted work cost with and without delegation? It may justify keeping the scout, narrowing its assignment or dropping it. More parallel sessions make sense after a checkable result has moved the checkout task forward at an acceptable total cost.