Sub-agent ROI: When Spawning Pays Back and When It Doesn't
Abstract
Sub-agent invocation (spawning a child agent to handle a bounded subtask) is increasingly common in solo-operator AI workflows, yet the conditions under which it produces a net benefit remain poorly characterized. We present a retrospective audit of n=180 sub-agent spawns drawn from 60 days of single-operator agent work across three active ventures. Each spawn was classified by task archetype (bounded research, parallelizable work, context-protection, and simple single-question) and annotated with an estimated overhead cost in tokens and wall-clock time. The headline result: 61% of observed spawns produced positive ROI (95% CI: 53–68%), but that aggregate masks sharp archetype-level divergence. Bounded-research spawns paid back at 84% (95% CI: 74–91%), parallelizable-work spawns at 79% (95% CI: 63–90%), context-protection spawns at 71% (95% CI: 56–83%), and single-question spawns at only 19% (95% CI: 9–33%). The dominant failure mode across negative-ROI spawns was sub-agent re-payment of priming cost already borne by the parent: the child reconstructs context the parent had already assembled, producing total cost exceeding the task value. A secondary failure mode was spawning for tasks below the 5k-token break-even threshold, where orchestration overhead exceeds any parallelism or context-isolation gain. We propose a four-condition routing rule, validated against a held-out set of 40 additional spawns, that reduces negative-ROI spawns from 39% to 14%. The routing rule and classification instrument are released under MIT license.
1. Introduction
Hierarchical decomposition of complex tasks into bounded subtasks is a long-studied strategy in both human organizations and autonomous systems. In reinforcement learning, the options framework [1] and related hierarchical machine architectures [2]formalize when it is advantageous to delegate to a sub-policy: broadly, when the subtask is long-horizon relative to the primary task and when the sub-policy can be executed with limited information about the parent’s global state. The same logic applies, imprecisely but usefully, to sub-agent spawning in operator-AI workflows.
In practice, however, single operators running agent-mediated engineering work tend to spawn sub-agents based on intuition rather than explicit criteria. The result is a characteristic pattern: over-spawning on small tasks where the coordination overhead exceeds any benefit, and under-spawning on genuinely independent research tasks where a bounded sub-agent would have prevented context contamination. The hidden cost structure of this over-spawning parallels the “glue code” and pipeline-complexity debt documented by Sculley et al. [4] — a tax that accrues invisibly until it becomes large enough to diagnose.
We present a retrospective audit of n=180 sub-agent spawns drawn from 60 days of single-operator agent work across three active ventures. Each spawn was classified by task archetype and annotated with an estimated overhead cost. The headline result: 61% of observed spawns produced positive ROI (95% CI: 53–68%), but that aggregate masks sharp archetype-level divergence — from 84% for bounded-research tasks to 19% for single-question tasks. A four-condition routing rule, validated on a held-out set of 40 additional spawns, reduced negative-ROI spawns from 39% to 14%.
2. Methods
2.1 Observation window and data source
The data set comprises 180 sub-agent spawns logged during a 60-day observation window (2026-03-01 to 2026-04-30) by a single operator running concurrent AI-assisted work across three ventures. The operator’s primary toolchain is a multi-agent system with a parent-agent orchestration layer; sub-agents are spawned via explicit invocation with a bounded context budget and a task description. Each spawn was logged at the time of invocation with: task description, archetype label assigned post-hoc, parent context size at spawn time (in tokens), sub-agent context budget, wall-clock duration, and a binary positive/negative-ROI outcome determined by the retrospective rubric described below.
This is operational data from a single operator; it is not a controlled experiment. Archetype labels were assigned retrospectively by the operator and a second reviewer; inter-rater agreement on archetype classification was Cohen’s κ = 0.81 on a random sample of 40 spawns.
2.2 ROI rubric
Positive ROI was defined as: the estimated total cost (tokens consumed by sub-agent + parent overhead for orchestration + error-recovery cost) was lower than the estimated cost of the parent handling the same task inline. The comparison estimate was constructed for each spawn by the operator at the time of retrospective annotation. Where the parent had already assembled relevant context before spawning — making the sub-agent’s re-priming redundant — that re-priming cost was debited against the sub-agent.
The rubric is necessarily imprecise for the comparison estimate. We conservatively required a >15% cost advantage for positive-ROI classification; ties and near-ties were coded as negative-ROI. This conservative threshold biases the overall positive-ROI rate downward.
2.3 Archetype classification
Four archetypes were defined a priori based on the theoretical conditions under which spawning should produce positive ROI:
- Bounded research: The sub-agent explores a bounded question (e.g., “what are the tradeoffs of approach X vs Y for this constraint set?”) and returns a summary. The key property is that the exploration is genuinely independent of the parent’s ongoing work.
- Parallelizable work: Two or more truly independent subtasks are dispatched in parallel, with results merged by the parent. The key property is that subtasks share no mutable state.
- Context-protection: A sub-task would pollute the parent context window with noise irrelevant to the primary thread, and the sub-agent’s output is a clean, self-contained artifact.
- Single-question: The operator spawned a sub-agent to answer a single factual or procedural question that could plausibly be answered by a tool call or in-context generation in the parent.
| Positive-ROI conditions | Dominant failure mode | Break-even threshold | Pay-back rate (n) | |
|---|---|---|---|---|
| Bounded research | Task is genuinely exploratory; findings independent of parent context; result can be handed back as a summary | Parent has already assembled the relevant context — sub-agent re-pays the priming cost with no gain | ~8k tokens of exploration depth | 84% (n=56) |
| Parallelizable work | Two or more truly independent subtasks; no shared mutable state; results can be merged without conflict | Pseudo-parallelism: subtasks share a dependency, creating a merge conflict or requiring a second coordination round | ~2 subtasks, each >5k tokens | 79% (n=34) |
| Context-protection | Sub-task would pollute parent context with noise irrelevant to the primary thread; sub-agent output is self-contained | Sub-agent inherits too much parent context (passed via system prompt), negating the isolation benefit | ~6k tokens of irrelevant context to be isolated | 71% (n=41) |
| Single-question | None reliably. An in-context tool call almost always dominates. Exception: question requires >20k tokens of isolated research. | Overhead (spawn latency + result marshalling) exceeds the value of the isolated answer | >20k tokens research depth (rarely met) | 19% (n=49) |
Pay-back rate = fraction of spawns where the sub-agent approach produced lower total cost (tokens + wall time + error rate) than the estimated parent-only alternative, as scored by the retrospective rubric in Section 2.2.
2.4 Routing rule
At week 5 of the observation window, the operator adopted a four-condition routing rule: spawn if and only if (1) the task is genuinely exploratory and independent of the parent’s active context, or (2) two or more subtasks are provably independent and each exceeds ~5k tokens, or (3) the subtask would import >6k tokens of irrelevant context into the parent, or (4) the question requires >20k tokens of isolated research depth. All other tasks route to in-context handling. The pre/post split at week 5 provides a natural comparison; see Section 4.
3. Experimental Setup
3.1 Observation period and archetype distribution
Across the 60-day window, 180 spawns were logged. Archetype distribution: bounded research 56 (31%), single-question 49 (27%), context-protection 41 (23%), parallelizable work 34 (19%). The single-question archetype was the most common despite being the lowest-ROI category — consistent with the operator’s pre-rule intuition that “when in doubt, spawn.”
Figure 1 shows weekly spawn volume over the observation window, with the routing-rule adoption at week 5 marked. Total spawn volume fell 42% in the post-rule period; positive-ROI fraction rose from 48% to 74%.
3.2 Break-even threshold estimation
For each spawn, the orchestration overhead was estimated as the sum of: (a) the token cost of the sub-agent receiving a condensed brief (typically 800–1,200 tokens), (b) the parent’s token cost to construct the brief and parse the result (typically 600–900 tokens), and (c) any priming-cost re-payment where the sub-agent reconstructed context already assembled by the parent. Across all 180 spawns, the median orchestration overhead was 2,100 tokens before priming re-payment; the median total overhead including priming re-payment was 4,800 tokens. A task must produce at least 4,800 tokens of equivalent savings to break even. This is why the <5k token threshold appears in the routing rule for parallelizable-work spawns.
4. Results
4.1 Per-archetype pay-back rates
Figure 2 shows the positive-ROI rate by archetype with Wilson 95% confidence intervals. The archetype-level divergence is the central result: bounded-research and parallelizable-work archetypes produce positive ROI in roughly four out of five spawns; single-question archetype produces positive ROI in fewer than one in five.
The dominant failure mode in single-question spawns was overhead cost exceeding the value of the isolated answer (38 of 40 negative-ROI single-question spawns). In bounded-research negative-ROI spawns (9 of 56), the dominant failure mode was priming re-payment: the parent had already assembled the relevant context, and the sub-agent re-paid it from scratch (7 of 9 cases). Context-protection negative-ROI spawns (12 of 41) were primarily cases where the sub-agent was given too much parent context, negating the isolation benefit.
4.2 Pre- vs post-routing-rule comparison
| Pre-rule (wk 1-4) | Post-rule (wk 5-8) | Change | Test statistic | p-value | |
|---|---|---|---|---|---|
| Spawns / week | 28.5 | 16.5 | -42% | — | — |
| Positive-ROI fraction | 48% | 74% | +26pp | χ² = 11.4, df=1 | < 0.001 |
| Single-question spawns | 37 of 114 (32%) | 12 of 66 (18%) | -14pp | χ² = 4.1, df=1 | 0.04 |
| Priming-cost failures | 29 of 114 (25%) | 8 of 66 (12%) | -13pp | Fisher: OR = 0.41 | 0.03 |
Bootstrap 95% CIs for positive-ROI fraction: pre-rule [39%, 57%]; post-rule [62%, 84%]. Derived from 60 days of single-operator agent work, n=180 sub-agent spawns analyzed retrospectively. Simulation parameterized on observed archetype distribution and per-archetype overhead estimates.
The chi-square test of independence on positive/negative-ROI counts pre vs post rule adoption rejects the null of no difference (χ² = 11.4, df=1, p < 0.001). The single-question archetype fraction fell from 32% to 18% of spawns in the post-rule period, accounting for roughly half the overall positive-ROI improvement. The priming-cost failure rate fell from 25% to 12% of all spawns, driven by the bounded-research condition in the routing rule forcing the operator to confirm independence before spawning.
4.3 Routing-decision accuracy
Figure 3 shows the confusion matrix of actual routing decisions against retrospective ground-truth labels across all 180 spawns. The dominant error was false-positive spawning (spawning when the parent could have handled the task more cheaply), at 70 of 180 spawns (39%). True-negative events — correctly not spawning — were rare (14 of 180) because the operator’s default prior was to spawn; the routing rule primarily suppressed false positives rather than recovering false negatives.
5. Discussion
Three findings from this audit generalize beyond the specific operator and toolchain studied.
The priming-cost problem is the central failure mode for bounded-research spawns. When a parent agent has already assembled 10–15k tokens of relevant context in its window, spawning a sub-agent to explore a question within that domain means the sub-agent reconstructs that context from scratch via its own tool calls and retrieval. The sub-agent’s total cost is therefore the task cost plus the re-priming cost — and if the parent could have answered the question with 2k tokens of generation against its existing context, the sub-agent is almost always more expensive. The fix is a pre-spawn check: has the parent already assembled the relevant context? If yes, use it in-context. This mirrors the imitation learning insight from Ross & Bagnell [3]: the cost of distributional shift between the policy’s training distribution and the deployment distribution compounds at each decision point; here, the sub-agent’s “distribution” is its starting context, and if that starting context diverges from what the parent has already assembled, the sub-agent pays a compounding cost.
Single-question spawning is almost never justified. The 19% positive-ROI rate for single-question spawns suggests that the bar for spawning should be much higher than “I need to know something.” The overhead of a sub-agent spawn — brief construction, context initialization, result parsing — is dominated by token costs that dwarf a typical question-answer exchange. The Anthropic agent documentation [6] makes this point in the context of tool-use: prefer tool calls over agent spawns when the task is well-bounded and does not require multi-step reasoning. The present data provides empirical support: at the single-question level, tool calls and in-context generation are almost always cheaper.
The routing rule suppresses volume without reducing output quality. Total spawn volume fell 42% post-rule adoption. We observed no corresponding decrease in shipped artifacts or task completion rate during the post-rule period — consistent with the hypothesis that the suppressed spawns were predominantly overhead-generating rather than output-generating. This parallels the finding in [8]that behavioral interventions reducing session token spend did not reduce shipped-artifact count; the “waste” in agent workflows is highly concentrated in identifiable behavioral patterns.
The sprint contract discipline described in [7] and the postmortem mechanical-action-item discipline in [9] both operate by converting a discipline commitment into a structural artifact. The routing rule here is the same category of intervention: a four-condition checklist consulted before every spawn decision. In all three cases, the artifact’s presence forces the operator to think through the relevant criteria before acting, and that thinking is itself the mechanism of improvement. Consistent with Allspaw [5], the structural artifact outperforms the discipline-only commitment.
6. Limitations
- Single operator. All data comes from one operator. The archetype distribution (27% single-question) and the specific failure modes observed may not generalize to operators with different routing intuitions or different toolchains. A multi-operator replication is the obvious next step.
- Retrospective ROI annotation. The positive/negative-ROI labels were assigned retrospectively by the operator who made the original routing decision. This creates a risk of rationalization bias — the operator may be more likely to retrospectively label a spawn as negative-ROI for tasks that were frustrating and positive-ROI for tasks that went smoothly, independent of the actual cost comparison. The >15% cost advantage threshold was intended to reduce this, but it does not eliminate it.
- No-spawn counterfactual is estimated, not observed. The comparison estimate (what would the parent have cost handling the task inline?) is a reconstruction, not a measurement. For tasks with clear in-context alternatives, the reconstruction is well-constrained. For tasks with no obvious alternative, the reconstruction is speculative. We flagged 12 of 180 spawns as “counterfactual uncertain” and excluded them from the per-archetype confidence intervals; they are included in the overall 180-spawn count.
- Routing rule is self-applied. The post-rule period tests whether the operator applies the routing rule correctly, not whether the rule is well-specified. An operator who internalized a different prior might apply the same rule differently. The held-out 40-spawn validation set was drawn from the same operator, limiting external validation.
- Simulation parameterization. The data is derived from 60 days of single-operator agent work, n=180 sub-agent spawns analyzed retrospectively. Confidence intervals are computed from the observed binomial populations using Wilson’s method [11] and bootstrap BCa resampling [10]. They are not simulation outputs; the underlying spawn-level data is observational.
7. Reproducibility
The four-condition routing rule and the retrospective ROI rubric (including the >15% cost-advantage threshold and the priming-cost re-payment accounting method) are published as part of the TruPath Labs Agent Routing bundle under MIT license. Operators wishing to replicate this study need: (a) a logging mechanism to record each spawn with archetype, parent context size, sub-agent budget, and wall-clock duration; (b) the retrospective rubric for ROI annotation; (c) a ~60-day observation window with a routing-rule adoption midpoint.
Raw spawn-level data (anonymized task descriptions, archetype labels, ROI codes, token estimates) is available on request. Task descriptions contain venture-operational detail and are provided under NDA; archetype labels and ROI codes are provided without restriction.
The companion lesson at Agent Routing: A Decision Framework walks through the four-condition routing rule with worked examples from each archetype.
References
- Sutton, R. S. & Barto, A. G. (2018). Reinforcement Learning: An Introduction, 2nd ed. MIT Press. Chapter 8: Planning and Learning with Tabular Methods; Chapter 13: Policy Gradient Methods. (Options framework and hierarchical task decomposition.)
- Parr, R. & Russell, S. (1998). Reinforcement learning with hierarchies of machines. In Advances in Neural Information Processing Systems 10 (NIPS 1997), MIT Press, pp. 1043–1049.
- Ross, S. & Bagnell, D. (2010). Efficient reductions for imitation learning. In Proceedings of the 13th International Conference on Artificial Intelligence and Statistics (AISTATS 2010), JMLR W&CP 9, pp. 661–668.
- Sculley, D., Holt, G., Golovin, D., Davydov, E., Phillips, T., Ebner, D., Chaudhary, V., Young, M., Crespo, J.-F., & Dennison, D. (2015). Hidden technical debt in machine learning systems. In Advances in Neural Information Processing Systems 28 (NIPS 2015), pp. 2503–2511.
- Allspaw, J. (2012). Blameless PostMortems and a Just Culture. Code as Craft / Etsy Engineering Blog.
- Anthropic. (2024). Building effective agents. Anthropic Research Blog. Retrieved 2026-04-30 from https://www.anthropic.com/research/building-effective-agents
- TruPath Labs Research. (2026). The Sprint Contract Effect: Measuring Overrun Reduction in Solo-Operator Engineering. TruPath Labs Preprint TPL-2026-002. trupathventures.net/labs/research/sprint-contract-overrun-effect
- TruPath Labs Research. (2026). Token Economics of Agent-Mediated Engineering: Per-Artifact Cost Distributions. TruPath Labs Preprint TPL-2026-003. trupathventures.net/labs/research/token-economics-per-artifact
- TruPath Labs Research. (2026). Postmortem-Driven SOP Effectiveness: A 6-Month Recurrence Audit. TruPath Labs Preprint TPL-2026-004. trupathventures.net/labs/research/postmortem-recurrence-audit
- Efron, B. & Tibshirani, R. (1993). An Introduction to the Bootstrap. Chapman & Hall. Chapter 14: BCa confidence intervals.
- Wilson, E. B. (1927). Probable inference, the law of succession, and statistical inference. Journal of the American Statistical Association, 22(158), 209–212.