- In short
- The high-leverage fix principle says to prefer the simplest, highest-impact change first when a Claude toolset misbehaves: better descriptions before scoped access, routing classifiers, or tool consolidation. It is an exam-favoured heuristic for evaluating competing solutions, rewarding the cheapest change that addresses the root cause over elaborate engineering.
What the high leverage fixes principle says
By the time you reach this knowledge point you can diagnose misrouting, split overloaded tools, write strong descriptions, and spot prompt conflicts. The high-leverage fix principle is the judgement that orders those moves. It states that when a toolset misbehaves, you should reach first for the change that delivers the most reliability for the least effort, and escalate to heavier solutions only when the cheaper one is genuinely insufficient. In tool design that means better descriptions before scoped access, scoped access before routing classifiers, and consolidation last of all.
This is an evaluate-level skill, the highest Bloom level in the task statement. The exam is not asking you to recall that descriptions matter; it is asking you to weigh several legitimate fixes against each other and pick the one with the best ratio of impact to cost. Two of the candidate answers will usually work. The credited one is the cheapest that addresses the actual cause.
- High-leverage fix principle
- A decision heuristic for Claude tool design: when selection or behaviour breaks, apply the simplest change that targets the root cause first, escalating through better descriptions, scoped access, routing classifiers, and tool consolidation only as each cheaper option proves insufficient.
The order of preference
The principle is concrete because it comes with a ranked ladder. From cheapest and most targeted to most expensive and structural:
- Better descriptions, rewrite or expand the tool descriptions with explicit boundaries. No new infrastructure, directly targets the selection signal, reversible in minutes. Almost always the right first move.
- Scoped access, give each agent only the tools its role needs, so the model is choosing among fewer, more relevant options. A configuration change, not a build.
- Routing classifiers, add a separate decision layer that pre-sorts requests to tools. A real system to design, test, and maintain; justified only when description and scoping fixes have failed.
- Tool consolidation, merge or restructure the tool surface itself. The most invasive, because it changes downstream handlers and contracts; a last resort when the toolset is fundamentally mis-shaped.
Each rung up the ladder costs more effort and adds more moving parts. The discipline is to climb only as far as the evidence forces you. Anthropic's own guidance echoes this: it frames detailed tool descriptions as the highest-leverage tuning surface and urges keeping agent designs as simple as the task allows, adding complexity only when it demonstrably improves outcomes.
Why cheaper-first is not just frugality
Preferring the low-effort fix is not a budget preference; it is a correctness heuristic. The cheapest fixes in this domain happen to be the ones that target the root cause most directly. Misrouting is usually caused by the descriptions, so editing the descriptions attacks the cause head-on. A routing classifier, by contrast, builds a compensating layer around the cause without removing it. You now maintain a second system whose only job is to paper over two sentences you could have rewritten. When you reach for heavy machinery first, the usual reason is that you misdiagnosed the cause, and the machinery then hides the misdiagnosis instead of fixing it.
There is also a compounding cost. Every classifier, every consolidation, every extra layer is something to test, monitor, and keep in sync as the toolset evolves. Anthropic's writing on advanced tool use treats techniques like tool search and consolidation as answers to genuinely large or unwieldy tool surfaces, not as default first responses to a routing bug. Simplicity bought early keeps the system legible; complexity added prematurely is paid for on every future change.
Evaluating a fix in practice
To apply the principle, ask two questions of any proposed fix. First, does it target the root cause, or does it compensate for it? Second, is it the cheapest option that does so? A fix that targets the cause but is expensive should still lose to a cheaper fix that also targets the cause. A cheap fix that only masks the symptom should lose to the targeted one even if the targeted one costs a little more. The credited answer is the intersection: cheapest and root-cause-targeting. That ranking is what lets you choose confidently among options that all "work."
Worked example
An agent with twelve tools frequently misroutes. Four proposals are on the table: rewrite the descriptions; scope tools to each agent's role; train a routing classifier; consolidate the twelve tools into five. You can do exactly one first.
Evaluate each against root cause and cost. Rewriting the descriptions is the cheapest change and targets the selection signal directly, if the misrouting comes from thin or overlapping descriptions, this fixes the cause in minutes and is trivially reversible. Scoping tools to roles is also cheap and reduces the number of options the model weighs, which helps, but it treats the volume of tools rather than the quality of their descriptions; useful, slightly less targeted to a pure misrouting symptom. A routing classifier is a separate system to build, evaluate, and maintain, and it compensates for ambiguity rather than removing it. Consolidating twelve tools into five restructures the whole surface and rewrites downstream handlers, the most invasive option by far.
Rank them: better descriptions first, because they are both the cheapest and the most direct attack on the likely cause. If, after rewriting, the descriptions are demonstrably clean and misrouting persists, climb one rung to scoped access. Only if scoping also fails do you justify a classifier, and only a fundamentally mis-shaped toolset justifies consolidation. The exam will offer the classifier or the consolidation as a sophisticated-sounding trap; the high-leverage answer is the description rewrite, because nothing cheaper addresses the cause and nothing that addresses the cause is cheaper.
Common misreadings to avoid
Misconception
A routing classifier or consolidation is the proper, professional fix for a misrouting agent.
What's actually true
Misconception
Preferring the simple fix is just about saving development time.
What's actually true
The principle beyond tool design
Although this knowledge point lives in the tool-design task statement, the high-leverage instinct is one of the most portable ideas on the whole exam, and recognising its reach helps you apply it confidently. The same ladder, cheapest and most targeted change first, heavier and structural changes only when forced, governs decisions across every domain. When an agent loses context, the first move is usually a better-structured prompt or a persistent facts block, not a new orchestration layer. When extraction quality dips, the first move is clearer criteria or a few well-chosen examples, not a bespoke validation pipeline. The surface details change; the ranking discipline does not.
Seeing the principle as domain-general changes how you read scenarios. Whenever a question offers a spread of fixes from a quick edit to a full re-architecture, you can apply the same evaluation: which option targets the root cause, and which of the targeting options is cheapest? The elaborate choice is nearly always a distractor designed to reward over-engineering. Training yourself to climb from the bottom of the ladder, in any domain, converts a large family of questions into a single repeatable judgement.
What leverage actually measures
It is worth being precise about the word leverage, because it is doing real work in the principle. Leverage is impact divided by effort, how much reliability a change buys per unit of work and ongoing maintenance. A description rewrite scores extremely high: minutes of effort, a direct attack on the cause, and nothing to maintain afterwards. A routing classifier scores low even when it works: substantial effort to build and evaluate, plus a permanent maintenance burden as the toolset evolves and the classifier must be kept in sync.
This is why the principle is not simply do less. A fix can be cheap and still wrong if it does not touch the cause; a few-shot example that masks misrouting is low effort but also low leverage, because the underlying ambiguity remains and will resurface. The highest-leverage option is the one that sits at the intersection of cheap and causally correct, and ranking candidates on both axes at once is the evaluative skill being measured. An architect who weighs only effort, or only impact, will pick wrong; the discipline is to weigh the ratio, and to prefer the change that delivers the most durable reliability for the least lasting complexity.
How it shows up on the exam
This is one of the most reliably tested heuristics in Domain 2, and it appears across scenarios, not just tool questions. You will be given a misbehaving system and four fixes spanning the ladder, and asked which to do first. The trap is that the elaborate options, build a classifier, consolidate the tools, add a machine-learning router, sound like the work of a serious engineer. The credited answer is almost always the description or scoping fix, because the exam rewards the cheapest change that addresses the cause. Internalise the ladder and you can answer a whole family of questions: when in doubt, climb from the bottom and stop as soon as the problem is solved.
A useful closing test is to ask, for any fix you are about to endorse, what you would have to maintain forever as a result of choosing it. A description rewrite leaves nothing behind to maintain; it simply makes the existing toolset work better. A classifier, a consolidation, or a new routing layer each adds a permanent obligation that must be kept correct as the system grows. Phrasing the choice as a question about lasting obligation, rather than just immediate effort, makes the high-leverage answer obvious and exposes the elaborate distractors for what they are: solutions that buy a marginal gain at the price of indefinite upkeep. That single question generalises the principle into a habit you can apply to almost any design decision the exam puts in front of you.
A multi-agent system misroutes because several agents share an overlapping toolset. An architect proposes building a routing classifier as the first intervention. Evaluating the options by the high-leverage principle, what is the strongest objection?
People also ask
What is the order of preference for fixing tool selection?
When is a routing classifier the right choice?
Why does the exam prefer simple fixes?
Watch and learn
Official Anthropic Academy lessons first, then hand-picked walkthroughs. Videos load only when you press play.
How We Build Effective Agents: Barry Zhang, Anthropic
Why watch: Anthropic's core 'keep it simple' and 'choose high-leverage' principles map directly to preferring simple description fixes before engineering complex routing solutions.
More videos for this concept
References & primary sources
Master this concept with Archie
Practice it inside an adaptive study session. Archie, your Socratic AI tutor, tracks your mastery with Bayesian Knowledge Tracing and schedules the perfect next review.