Which agent scaffolding survives a generation jump — few-shot dies, decomposition doesn't, verification splits in half
The question
"Which categories of agent capability scaffolding (few-shot examples, task decomposition, verification steps) empirically survive multiple frontier-model generation jumps, versus needing to be rebuilt each time?"
This sharpens the open question left unanswered in [[2026-08-01-scaffolding-vs-harness-the-erosion-axis]] — whether category-1 capability scaffolding erodes uniformly, or whether some kinds outlive others. The note's stated guess was that verification persists longest. That guess is half right, and the half that is wrong is the half we were about to sell to a client.
What we already know (from the vault)
- The three-category frame is already ours. [[2026-08-01-scaffolding-vs-harness-the-erosion-axis]] splits scaffolding into capability scaffolding (erodes), particular knowledge (compounds), and taste/policy (compounds), and files decomposition, verification, and few-shot all together in the eroding bucket. This brief keeps the frame and re-sorts the bucket.
- First-party guidance already killed few-shot, explicitly. [[2026-07-24-thariq-context-engineering-claude-5-rules]] — Anthropic deleted roughly 80% of Claude Code's system prompt for the Claude 5 generation with no measurable coding-eval loss, and rule 2 of the then/now table is "Give Claude examples → Design interfaces," on the reasoning that examples constrain the exploration space where an expressive tool parameter does the same job without the constraint.
- We have one real longitudinal breakage on file. [[2026-07-25-every-opus5-compound-engineering-breakage]] — Every's Compound Engineering plugin, built for the Opus 4 generation, broke on Opus 5: premature turn-end at phase seams, the agent returning control mid-autonomous-flow. The remediation was a 21% corpus deletion. The single most transferable rule they extracted is a verification rule: verify a phase ran by artifact existence, never by the final message — learned after an over-cut run skipped two whole phases and reported success with a clean transcript.
- Our own critic stack is mostly the fragile kind. [[2026-07-30-multi-llm-critic-ensembles-differentiated-rubrics]] establishes that
station-criticfans out one subagent per rubric axis. That is model-checking-model verification with an external rubric bolted on — and per the finding below, the rubric is the durable part, not the ensembling. - Optimizing a skill against a held-out score is already in the vault as a pattern, not a build: [[2026-05-26-skillopt-self-evolving-agent-skills]] treats a skill doc as frozen weights and accepts an edit only on strict held-out improvement.
What the web says
- Few-shot exemplars stop functioning as reasoning aids and survive only as format hints. "Revisiting Chain-of-Thought Prompting: Zero-shot Can Be Stronger than Few-shot" (arXiv 2506.14641) reports Qwen2.5-72B at 95.83% zero-shot vs 95.75% 8-shot on GSM8K, and 81.64% vs 81.30% on MATH — inside noise. The load-bearing result is the noise ablation: replacing 50% of exemplar tokens with "XXX", or shuffling word order entirely, leaves strong-model accuracy stable or slightly better. Attention maps show the exemplar block getting low attention. Older/weaker models (LLaMA2-7B) still gain measurably. The authors' mechanism: exemplars align output format, and once evaluation was corrected to read
\boxed{}answers, most of the apparent few-shot advantage evaporated. - Task decomposition scaffolding did NOT lose value on stronger models — the best controlled study says the opposite. "Scaffold Effects on GAIA: A Controlled Comparison" (arXiv 2606.08529) ran three scaffolds (ReAct; Planner-Actor-Rater; Planner-then-Executor) across Claude Opus 4.7, Sonnet 4.6, Haiku 4.5, Gemini 3.1 Pro Preview and GPT-5.5 on GAIA L1/L2 — 6,255 designed runs, 5,907 clean. Max-minus-min accuracy spread across scaffolds at L2 (robust slice): Opus 27.9pp, Haiku 26.2pp, Sonnet 22.1pp, Gemini 16.7pp, GPT-5.5 5.8pp. The most capable Anthropic model was the most scaffold-sensitive, not the least. The authors state the shrinking-sensitivity hypothesis was contradicted.
- The same study shows decomposition buying efficiency and error recovery, not just raw score. At L2, ReAct averaged 22.0 tool calls and 4.8k output tokens; Planner-then-Executor 6.4 calls and 2.9k tokens. In-trajectory error recovery went from 38% (ReAct) to 57% (Planner-Actor-Rater) / 51% (Planner-then-Executor). Those are the kinds of gains a smarter model doesn't automatically absorb.
- Model-internal verification is dying on a measurable curve. "Self-Consistency Is Losing Its Edge" (arXiv 2511.00751) reports 0.4% accuracy gain on HotpotQA across 20 sampled paths and 1.6% on MATH-500; Gemini-2.5-Pro on MATH-500 goes 98% → 99.6% for roughly 15x the token cost, with plateaus and occasional declines at high sample counts. The stated mechanism is the whole story: aggregation feeds on disagreement between samples, and a stronger model produces less disagreement. Best-of-N is separately bounded by verifier precision, so it plateaus wherever no automatic verifier exists.
- Scaffolding effort has historically bought about a third of what a generation jump buys. METR's post-training-enhancement measurement (metr.org) found OpenAI's own post-training moved GPT-4 agent success from 5% to 30% on a 195-task suite, while METR's full enhancement stack (better prompting/chat format, reasoning + file-edit + browse + vision tools, reward-model selection over 8 candidate actions, light fine-tuning) added 30% → 38% — 8pp, not statistically significant at ±6.
- The GAIA and METR numbers are not in conflict, and the reconciliation is the point. METR measured incremental tooling piled on an already-decent scaffold and got a noisy 8pp. GAIA measured choice of scaffold architecture and got a 28pp spread on a frontier model. The productive cut is not how much scaffolding, it is which shape.
Convergences and contradictions
- Convergence on few-shot, from two independent directions. The academic ablation (models literally ignore corrupted exemplars) and Anthropic's own product decision (delete examples, design interfaces instead) reach the same place by different routes. This is the one category where "rebuild each generation" is the wrong answer — the right answer is stop building it, and get output shape from a schema or an expressive tool parameter, which costs less and doesn't rot.
- Direct contradiction with our own concept note on decomposition. The erosion-axis note files decomposition under "decompose because it can't plan" — pure capability crutch, expected to erode. The GAIA controlled comparison says scaffold sensitivity rose rather than fell with capability inside the Anthropic family, and that planner/executor structure cut tool calls ~3.4x while raising error recovery ~13-19pp. Our note was reasoning from the bitter lesson; the measurement disagrees.
- Partial contradiction with our own note on verification. The note guessed verification persists longest. The evidence says verification is not one category. Self-consistency and same-model best-of-N erode faster than few-shot, because their input is literally model error and the supply of that is shrinking. Artifact-existence and test-pass verification show no decay mechanism at all. Every's post-mortem is the receipt: the rule that survived an actual generation jump was the artifact-existence one.
- One clean unresolved contradiction: GPT-5.5's 5.8pp spread. It is by far the least scaffold-sensitive model in the GAIA set, on identical tools and tasks — but Haiku 4.5, the weakest model tested, sits at 26.2pp, between Sonnet and Opus. So the ordering is not capability. It looks like a model-family / agentic-post-training property. Nobody has isolated it.
Synthesis for RDCO
The operating test in the erosion-axis note — is this here because the model is dumb, or because the model can't know this? — mis-sorts two of the three categories in this question. Decomposition and artifact verification are both there "because the model is dumb," by that phrasing, and neither eroded. The test needs replacing, and the replacement is a better question anyway: what is this scaffolding's input?
Scaffolding whose input is the model's own output distribution erodes as that distribution tightens. Few-shot exemplars shape token continuations; self-consistency harvests sample variance; a same-family critic reads a same-family producer. All three get their leverage from model error, and every generation jump takes some of that supply away. This category should be built cheap, instrumented, and deleted on schedule — and self-consistency specifically answers the note's second open question about a cost-driven fourth category: it is the fourth category, eroding on the capability curve and the price curve simultaneously, which is why it is the first thing to cut.
Scaffolding whose input is outside the model does not erode, because a smarter model does not change the input. A filesystem either has the plan file or it doesn't. A test either passes or it doesn't. KwikTrip's hot-dog-roller failure modes are either written down or they aren't. This is the same set the note already called particular knowledge, extended with a category it had misfiled — external-ground-truth verification. Decomposition sits in a third position: its input is the task's structure, not the model's, which is why GAIA found its value holding or rising with capability. On the note's own harness/scaffolding line, decomposition belongs closer to harness than to scaffolding, and re-filing it there is the cleanest fix to the framework.
For the KwikTrip conversation, this changes what we tell them to harden. The call on the document-preprocessing pipeline stands — that is genuine capability scaffolding against weak multimodal diagram retrieval, and it is a wasting asset. But the note as written implies the orchestration around Sprocket is disposable too, and we should not say that onsite. The planner/executor split and any artifact-level phase verification are the parts to harden, and they are unusually easy to defend to a client because their benefit is measurable in operational units — tool calls per resolved ticket, percentage of failed trajectories that self-recover — rather than in eval points that a model release can erase. What we should tell them explicitly not to harden: exemplar libraries of "good" answers, and any self-review pass where the reviewing model is the same model that produced the work.
The mirror is uncomfortable. Our own critic stack is largely the fragile kind: station-critic and the verify-* family are model-checking-model. The differentiated-rubric work is what rescues them, because a rubric is an external anchor — but the ensembling itself buys less every release, on exactly the self-consistency curve. Meanwhile the durable kind is missing where we know it should be: the Opus-5 breakage note already flagged that our rewritten deep-research skill asserts "<N> briefs filed overnight" with nothing checking that a brief file exists. That is the one verification category with no measured decay, and we do not have it. The honest read is that our recent skills-retune spent effort on the category that depreciates and skipped the category that doesn't.
Calibration on the evidence itself, plainly. No published study measures the same scaffold across successive generations of one model lineage. GAIA is cross-sectional — five models at one moment, capability tiers standing in for generation jumps — and the GPT-5.5 outlier is direct evidence that the stand-in is leaky. The few-shot ablation runs on open-weight models topping out at Qwen2.5-72B, not on frontier closed models; the frontier corroboration is Anthropic's product decision, which is first-party but not a controlled experiment. METR's number is 2024-era and its non-significant 8pp is a weak instrument. The self-consistency result is the cleanest of the four and it is also the narrowest. So: high confidence on the few-shot and self-consistency verdicts, where mechanism and measurement agree. Moderate confidence on decomposition — one good study, contradicting our prior, on a benchmark that rewards multi-step tool use. Low confidence on any specific decay rate for anything. Nobody should quote a half-life.
Why this is in the vault
It resolves the first of the two open questions in [[2026-08-01-scaffolding-vs-harness-the-erosion-axis]] and corrects that note's decomposition filing, which is the note we are drawing on for the KwikTrip Sprocket harden-vs-disposable conversation — specifically the recommendation on whether the client should treat orchestration structure as a durable asset alongside the knowledge base. It also names a concrete gap in RDCO's own skill corpus: artifact-existence verification is the one category with no measured decay, and deep-research currently reports brief counts without it.
Open follow-ups
- Is scaffold sensitivity a capability property or a vendor post-training property? GPT-5.5 showed a 5.8pp scaffold spread on GAIA L2 against Opus 4.7's 27.9pp on identical tools and tasks, while the weakest model tested sat in between — so the ordering is not capability. If sensitivity tracks how much agentic post-training a lab did, then scaffolding durability is a per-vendor question, and an Anthropic-only stack is structurally the most scaffold-dependent one on the market. Researchable against model cards, agentic post-training disclosures, and any cross-vendor scaffold comparison.
- Can model output variance be measured as a leading indicator for retiring scaffolding before the next release, rather than after? The self-consistency decay has an explicit mechanism — benefit is proportional to inter-sample disagreement. If that generalizes to every technique whose input is the model's own output distribution, then sampling entropy on our own live workloads is a retirement signal we could read now, instead of discovering erosion the way Every did, at the seam of a breaking generation jump.
Related
- [[2026-08-01-scaffolding-vs-harness-the-erosion-axis]]
- [[2026-07-24-thariq-context-engineering-claude-5-rules]]
- [[2026-07-25-every-opus5-compound-engineering-breakage]]
- [[2026-05-10-harness-moat-two-layers-portability]]
- [[2026-07-30-multi-llm-critic-ensembles-differentiated-rubrics]]
- [[2026-05-26-skillopt-self-evolving-agent-skills]]
- [[2026-07-31-kwiktrip-onsite-strategic-prep]]
- [[2026-05-01-claw-vs-harness-engineering]]
Sources
Vault
- 06-reference/concepts/2026-08-01-scaffolding-vs-harness-the-erosion-axis.md
- 06-reference/2026-07-24-thariq-context-engineering-claude-5-rules.md
- 06-reference/2026-07-25-every-opus5-compound-engineering-breakage.md
- 06-reference/concepts/2026-05-10-harness-moat-two-layers-portability.md
- 06-reference/research/2026-07-30-multi-llm-critic-ensembles-differentiated-rubrics.md
- 06-reference/2026-05-26-skillopt-self-evolving-agent-skills.md
- 06-reference/concepts/2026-05-01-claw-vs-harness-engineering.md
Web
- Starace et al., "Scaffold Effects on GAIA: A Controlled Comparison" — https://arxiv.org/html/2606.08529
- "Revisiting Chain-of-Thought Prompting: Zero-shot Can Be Stronger than Few-shot" — https://arxiv.org/html/2506.14641
- "Self-Consistency Is Losing Its Edge: Diminishing Returns and Rising Costs in Modern LLMs" — https://arxiv.org/html/2511.00751
- METR, "Measuring the impact of post-training enhancements" — https://metr.org/blog/2024-03-15-measuring-post-impact-enhancements/