Depth is sold as cheaper context. Multi-tier agents keep the leaves busy, push a shorter summary to the root, and in theory give you more working memory per dollar.
That pitch is half right. It confuses three things that do not move together: how many findings survive to the root, how much state the root has to carry, and what each run actually costs. Once you separate them, the trade looks less like "cheaper context" and more like a depth budget you have to re-derive against your own workload.
This is an operator read of a recent v1 preprint, not a benchmark recap. The paper pins down a retention model and a few production estimates. The three-axis framing and the per-tier alignment check are mine — they are how I would actually use the paper, and they are not findings I am attributing to the author.[^arxiv-v1]
Three axes, not one
Most agent conversations treat depth as a single dial. More tiers, more parallelism, smaller root context, lower cost per task. That intuition is the failure mode the preprint exposes.[^arxiv-v1]
In practice, depth moves three axes at once, and they are not the same axis:
- Finding yield at the root — how many of the items the leaves saw actually reach the final report.[^arxiv-v1]
- Root-context exposure — how much state the root has to keep across the run.[^arxiv-v1]
- Dollars per finding — the actual billing curve, not the theoretical one.[^arxiv-v1]
You can gain on one and lose on another. A second tier often shrinks root-context exposure while quietly raising dollars per finding through alignment losses and a longer tool loop. A flat agent often keeps a high yield while running root context up against its limit. The interesting question is never "is depth cheaper?" It is "on which axis, by how much, at the cost of which other axis?"
That is the operator move the rest of this post makes: stop optimizing depth. Start measuring it.
What the retention model actually says
The author proposes a retention model so the depth trade becomes analyzable, not magical.[^arxiv-v1] An agent handed b items keeps any one of them with probability r(b) = C · b^-δ. Stack k tiers over N findings and the tree yields roughly C^k · N^(1-δ).
Two things follow from the shape of that equation.
First, the task-size exponent δ survives any arrangement of agents. If δ is 0.34 on production traces — the author's estimate from 600 deep-research traces with a 95% confidence interval of [0.30, 0.38] across three identifications — no agent topology escapes that exponent.[^arxiv-v1] The leaves still drop items at the same rate, and the depth-k factorization only multiplies the constant C per tier, not the exponent. In the author's retention model, flat stays optimal for yield.
Second, the constant C is observed, not assumed. Across 16,082 tool-bounded hops where b = 1 occurred 550 times, the author reports C = 0.571 with a 95% confidence interval of [0.527, 0.615].[^arxiv-v1] That is the per-tier multiplier on findings that survive before any alignment loss is layered on top. It is not a universal property of every hop structure — it is a measured constant on this hop sample.
So the paper's first move is to remove depth as a free improvement on yield. The retention exponent is the floor under any multi-tier arrangement, and the constant C is what each tier compounds.
The empirical estimates the author reports
The numbers in the previous paragraph are the paper's, and they are worth holding to scope. Three more measurements make the rest of the argument concrete.[^arxiv-v1]
- Per-tier alignment: on 1,012 annotated multi-agent traces, one brief in sixteen went off-target, giving a per-tier alignment factor μ = 0.939 and a per-tier effective constant C·μ = 0.536. The author's reported value, not a deployment-ready threshold.[^arxiv-v1]
- Production cost scaling: in the paper's production-trace scope, a flat agent bills as N^1.39, not the N^2 an append-only context would predict. The 1.39 exponent is the author's reported scaling — not a universal billing law.[^arxiv-v1]
- Equal-spend crossover: at the paper's modeled cost curve and the reported C, μ, and 1.39 exponent, two tiers overtake flat at 403 findings under equal spend. The author's reported crossover under the paper's setup — not a threshold anyone should paste into their own billing.[^arxiv-v1]
Each of these numbers has a scope the article keeps adjacent to it. The 600 traces, the 16,082 hops, the 1,012 annotated briefs, the 403-finding crossover — none of them are universal. They are the scale on which the paper's retention model and cost model currently sit, and the right operator move is to re-derive against the local workload instead of transplanting the constants.
Why the root context is special
Depth does one thing the leaves cannot: it reduces what the root has to carry. In the author's model, root-context exposure drops from N items at depth 1 to roughly N^(1/k) at depth k. For depth 2 over 10,000 findings the modeled reduction is 10,000 items to 100 items.[^arxiv-v1]
The word "modeled" matters. That reduction is a derivation about how retention propagates through the tree, not an empirically proven persistence guarantee. There is no production trace measurement here that confirms "root context really carries 100 items at depth 2 over 10,000 findings." The author is reasoning about a clean abstraction.
I keep that distinction in mind because the abstraction is genuinely useful for sizing depth, and the abstraction is genuinely silent on real persistence. The operator move is to credit the reduction when reasoning about budget and to refuse to attribute a persistence guarantee the paper does not claim.
That alone is enough to justify treating root-context exposure as its own axis. It is the only place the tree's structure shows up at all.
The cost crossover is a curve, not a verdict
The 403-finding crossover is the number most readers will quote first. It is also the number most likely to be misused.[^arxiv-v1]
The crossover is the paper's reported point where two tiers become cheaper than flat at equal spend, given the author's N^1.39 cost curve, the reported C = 0.571, and the per-tier alignment μ = 0.939.[^arxiv-v1] Three things change the answer for any other workload:
- A different N^α production curve. The 1.39 exponent is the paper's measurement, not a law.[^arxiv-v1]
- A different alignment factor. Off-target rates in production traces will not match 1 in 16.
- A different retention exponent δ on the operator's own task distribution.
The operator move is to recompute the crossover against the local curves, not to import 403 as a deployment threshold. The general form the paper implies is unsatisfying but honest: two tiers beat flat at equal spend when the per-finding cost under depth k falls below roughly 1 / (C_local · μ_local)^k of the flat per-finding cost. Plug in the local numbers and the crossover moves.
The cost axis is the one where it is easiest to fool yourself. N^1.39 looks like a small exponent until the retention and alignment losses pile up at scale.
Delegation is an opening move
The paper's hazard-model result on 743,819 production tool calls is the most behaviorally interesting claim, and it sits on a different axis again.[^arxiv-v1]
The author finds that delegation does not respond to a filling context. It is an opening move, not a relief valve.
The fleet-instrumentation consequence is direct. If delegation is an opening move, the fill-and-delegate pattern that shows up in monitoring — many low-context handoffs early in the run, fewer late — is the agent doing what the author describes, not an artifact of context-window panic. The right operator reaction is to instrument opening delegation as a property of the harness, not as an emergency response.
The author's 7.8% delegation share in the production sample sits inside the model's predicted 0.7% to 11.3% range on the same data.[^arxiv-v1] That is a narrow window, not a free parameter. It does not generalize into a behavioral law across every agent system — the paper reports the result on its own 743,819 tool calls, and the number is what the operator reframes, not what the paper extrapolates.
A depth budget and an alignment check
The three axes and the per-tier number together give a usable budget — and this is where the post stops being a paper recap and becomes operator synthesis.
The three-axis depth budget (mine, not the paper's): treat yield retained, root-context exposure, and dollars per finding as three independent axes when sizing depth. Do not assume movement on one implies movement on the others. Re-derive thresholds against the local workload.[^arxiv-v1]
For yield at depth k, the budget form is C_local^k · N^(1 - δ_local). For root-context exposure, the budget form is N^(1/k) in the author's modeled reduction. For cost, the budget form is the operator's local N^α curve and the local billing model. Each axis is sized on its own scale and on its own measurement.
The per-tier alignment check (also mine): compute C_local · μ_local from the operator's locally measured per-tier constant and per-tier alignment. If C_local · μ_local falls below the paper's reported 0.536, the operator-side rule is to fall back to flat and re-measure rather than adding the tier.[^arxiv-v1] The paper reports the C, μ, and C·μ values; the gate is mine. Do not present C·μ = 0.536 as a deployment-ready threshold.
That gate is the useful artifact from this paper for an operator harness. The paper exposes the constants you would otherwise have to invent. The synthesis here is the promotion rule around those constants: a single threshold, named, before another tier is added to the tree.
A worked example, replayable against your own curves:
Re-derive (operator side): assume depth 2 over N findings. Use the author's constant C = 0.571 and exponent δ = 0.34 as anchor numbers, then replace them with C_local and δ_local from the operator's own traces. Yield reaching the root under the author's factorization is roughly 0.326 · N^0.66 for the paper's anchor values; the operator's workload should plug in C_local^2 and (1 - δ_local) on the same form. Root-context exposure, in the author's modeled reduction, is N^(1/2) — 100 items for 10,000 findings, not an empirical persistence claim. Cost crossover at equal spend uses the paper's N^1.39 curve: two tiers beat flat when per-finding cost under depth 2 falls below 1 / (C_local · μ_local)^2 of the flat per-finding cost, with the paper's anchor giving the 403-finding crossover you would re-derive, not transplant.
That worked example is the re-derivable artifact. A reader should be able to take C_local, δ_local, μ_local, and the local cost curve, plug them into the same three formulas, and get a depth decision without re-reading the paper.
What this changes for an operator harness
Three practical moves follow.
First, stop optimizing depth as a single dial. Track yield, root-context size, and dollars per finding on three curves. The retention model gives a clean factorization for yield; the cost model gives a clean curve for spend; the alignment check gives a clean gate for promoting another tier. The three live on separate graphs.
Second, instrument opening delegation. The hazard-model result says delegation is an opening move, not a response to a filling context. Treat it as a property of the harness the same way you would treat an initial tool selection: budgeted, logged, and reviewed.
Third, label the depth decision clearly. The three-axis budget and the per-tier C·μ alignment check are mine. The retention factorization, the 600-trace δ estimate, the 16,082-hop C observation, the 1,012-trace μ, the N^1.39 cost scaling, the 403-finding crossover, and the 743,819-tool-call hazard model are the author's. Keeping those labels adjacent to the claims they qualify is what stops this post from sliding into a "depth is broken" hot take or a paper recap. The value here is the operator framework around a v1 preprint, not the preprint itself.
That framing is also why this is a draft. The author has not peer-reviewed the result, and no independent replication is in evidence.[^arxiv-v1] The estimates are conditional on the retention assumptions, the task-size exponent, and the alignment losses. Treat the depth budget and the alignment check as operator-side templates to re-derive, not as rules to transplant.
The work to do next is the operator's: take the same three axes, plug in the local curves, and see where the crossover actually sits.
[^arxiv-v1]: He, R. (2026). Decomposition Buys Integrity, Not Yield. arXiv preprint arXiv:2609.17464v1. Cited as a v1 preprint under review; figures and modeling claims are bound to the v1 record at https://arxiv.org/abs/2609.17464 and may shift in later revisions.