--- title: Tuning agent prompts matters more than model choice,... description: A Salesforce AI research paper finds that once an agent's system prompt, tools, and workflow are carefully tuned around a weaker model, swapping in a stronger... canonical: https://daily.dev/posts/tuning-agent-prompts-matters-more-than-model-choice-salesforce-research-finds-k50dlsbqn twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Tuning agent prompts matters more than model choice, Salesforce research finds | daily.dev og:description: A Salesforce AI research paper finds that once an agent's system prompt, tools, and workflow are carefully tuned around a weaker model, swapping in a stronger... og:url: https://daily.dev/posts/tuning-agent-prompts-matters-more-than-model-choice-salesforce-research-finds-k50dlsbqn og:image: https://api.daily.dev/og/posts/k50dlsbQN.png og:image:alt: Tuning agent prompts matters more than model choice, Salesforce research finds og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Tuning agent prompts matters more than model choice, Salesforce research finds **[Collections](https://daily.dev/sources/collections)** · 5 min read · 1 upvotes · 0 comments ## Summary A Salesforce AI research paper finds that once an agent's system prompt, tools, and workflow are carefully tuned around a weaker model, swapping in a stronger model doesn't automatically improve results. The harness — prompt structure, tool definitions, and workflow logic — carries significant weight in agent performance, sometimes more than model choice itself. The practical implication is that teams building or benchmarking agentic systems should treat prompt optimization as essential and ensure the harness is tuned per-model before comparing model capabilities, since otherwise they're measuring prompt fit rather than true capability differences. ## Content Something has shifted in how I think about agents. The frontier models have clustered in capability, and the models release every few weeks anyway, so naming and benchmark bragging matter less than cost, tool reliability, and infrastructure efficiency. The model is one component inside a larger system with memory, tools, browsers, code execution, and access to enterprise systems. If you want an edge, it's in the harness. So here's my main advice: own your harness. ## System One and System Two I've started calling fast classification-style decision models System One models, and the big reasoning models System Two. Jev is the one I've been using most, and I'm seeing the difference in my own harnesses. Part of why this feels natural to me is history. I spent my entire PhD building efficient classifiers from scratch, from graph-based methods to deep learning, so watching classification models come back in the era of agents is a bit nostalgic. The emotion dataset we built at @dair_ai was used to test a new model, Julia-1, which was fun to see. It's also a good example of what open source enables. Julia-1 is still a work in progress, and the cloud GPU spend on training and experiments was about R$540 (US$104.08). I wonder whether we'll see a new kind of frontier lab focused on models like these. The reason I like this pairing is that a System Two model usually controls every component and decision in a harness, including many it doesn't need to. Hybrid systems have historically done better on domain-specific problems, and that still seems true with general-purpose frontier models. You're probably paying a premium for reasoning you don't need on routing, guardrails, and verification. So far I've had good results with guardrails, routing, and verifiers. Other places I'm starting to see promise: - structuring and optimizing skills - more efficient tool calling - dynamic workflows, where an orchestrator and subagents generate harnesses on the fly and a Jev-like model structures what passes between them - coordination between agents - smarter, faster loading of tools, MCP servers, and system prompts, and more broadly any part of context engineering that benefits from quick decisions - combining recursive language models with System One models There's a good post that tests Jev as a fuzzy linter, and it's worth reading before you write the idea off. It scopes the problem carefully: it filters down to rules that are easy to judge without extra context or heavy reasoning. That matters because System One models aren't built for reasoning-heavy work. They do help with the more deterministic decisions, if you look closely at what your harness is doing. I hope we get more evals and benchmarks for this, because there's a thread worth pulling. It's also encouraging that DSPy now supports System One models for calibrating agentic workflows. @lateinteraction relayed the framing from @typesafeai: DSPy methodology plus System One, program, don't prompt. I'm still surprised at how naturally these models slot into those frameworks. If you want to try building your own, my guide on a custom harness with Pi and Jev is a starting point. ## Keep it minimal There's a countertrend that I think matters just as much. Developers keep stacking complexity on top of agentic coding tools: skills, instructions, custom loops, agents, MCPs, memory systems, scripts. Then the harnesses and models improve, and the layer becomes unnecessary. @housecor put it well: often the extra complexity is actively harmful, burning more tokens for worse results. A paper from MIT CSAIL makes the same case for a deliberately minimal harness, and it's worth bookmarking. Another paper shows that the harness changes a coding agent's results by a lot, so harnesses do matter, but the answer isn't more of them. A separate paper shows how effective tuning an agent's system prompt can be, and it's recommended if you tune harnesses. New Salesforce research points the same way: once you tune prompts, tools, and workflow around a weaker model, copying a stronger model's setup stops being the obvious move. Some things do last. @geoffreylitt finds that agent skills can stay relevant across model generations as long as they're closer to specs than step-by-step instructions, for example a good format for a document output or a good flow for interviewing a human. @trq212 points out that you can barely "show your prompt" anymore, because the real work lives in references, skills, and examples. He often has his agent read three other repos he's made, search the web for references, and call other AI APIs first. @svpino likes an approach where you refine how your agent works by telling it how to behave. It works with any harness through an adapter, and he assumes the evolution lives in the adapter, so you can improve it over time and personalize it. ## What I'd do Build your harness, write the evals, and run the experiments yourself, especially ones that combine System One and System Two models. You'll be surprised how many problems a good classifier can solve, and how much of your frontier-model spend was for decisions that never needed it. Keep the scaffolding light, prefer specs over scripts, and delete layers when the models catch up. There are more layers to this, and I expect frontier labs to push decision models further and to ship better tools for customizing them. For builders who care about custom harnesses, this is a good moment. ## Questions this post answers ### Does upgrading to a stronger LLM automatically improve an AI agent's performance? Not necessarily. Salesforce AI research found that once an agent's system prompt, tools, and workflow are carefully tuned around a weaker model, simply swapping in a stronger model does not automatically yield better results. The harness—prompt structure, tool definitions, and workflow logic—carries substantial weight in overall performance, sometimes more than the underlying model's raw capability. _daily.dev surfaces research like this for engineers deciding whether to upgrade models or retune their agent harness._ ### Why might a benchmark comparing different LLMs inside an agent loop be misleading? Because if the prompt harness isn't separately tuned for each model, the benchmark ends up measuring prompt fit rather than true model capability. Salesforce research emphasizes that system prompts, tool definitions, and workflow logic are load-bearing components of agent performance, so an untuned harness can unfairly favor whichever model the prompt happened to be optimized for. _Teams comparing agent models track findings like this on daily.dev before trusting a benchmark._ ## Community take How the wider developer community reacted, aggregated from 1 discussion and 87 comments across x (as of 2026-09-30). **TL;DR:** Developers largely agree that custom scaffolding, skills, and multi-agent setups have a short shelf life as models improve, though many argue durable layers (verification, permission boundaries, project-specific facts) still earn their keep. **Sentiment:** 15% positive · 60% mixed · 25% skeptical **The case for** - Custom skills and scaffolding can still meaningfully improve output quality for people with deep domain expertise. - Layers that check work (running the app, browser passes) or encode stable project facts tend to survive model upgrades better than 'think harder' instructions. - Treating scaffolding as disposable and regularly deleting unneeded layers keeps setups lean and effective. **The pushback** - Heavy middleware and prompt scaffolding often become bottlenecks or actively restrict a smarter model's own reasoning. - Maintaining scaffolding after a model has absorbed the workflow wastes time and tokens, especially with stale memory notes. - Some see the whole trend of complex harness-building as premature abstraction or even a grift not backed by real data. **By community** - x (mixed): Replies split between people who still value hand-tuned skills/scaffolding for their domain and many others who say most custom layers get obsoleted by each model upgrade and should be deleted. **Hottest debate:** Whether custom skills, MCPs, and multi-agent scaffolding remain valuable additions or are mostly temporary crutches that smarter models render obsolete. **Open questions** - How do teams reliably decide which layers 'earn their keep' versus which are just patching a model's current weakness? - Is there objective data behind claims that skills/scaffolding help or hurt, beyond individual anecdotes? **Highlights** > @housecor The failure mode we see running production agents is treating prompt scaffolding as permanent architecture instead of temporary bridge work. When the underlying model capability jumps, heavy middleware actually restricts the model from reasoning through tool selection directly. > — [JimCen37 on x](https://x.com/JimCen37/status/2105052571585159313) > @housecor the layers that have survived model upgrades for me are the ones that check the work, like actually running the app and doing a browser pass. the ones telling the model how to think are what rot > — [shahidcodes on x](https://x.com/shahidcodes/status/2105036504536088982) > @housecor useful test: does the layer patch a model weakness or encode a fact about your project? weakness patches rot with every model upgrade. repo conventions, test commands and deploy rules keep paying off > — [OwaisLogs on x](https://x.com/OwaisLogs/status/2105053070178853035) > @housecor This is probably the biggest trap with agentic tooling: optimizing the harness faster than the models evolve. A workflow that gives you an edge today can become unnecessary—or actively harmful—after the next model upgrade. The best agent setup might be the one you’re willing to > — [09x88hq on x](https://x.com/09x88hq/status/2104989193025855985) > @sirstripy @housecor We have a similar experience And it is ridiculous. Along with MCPs, working in mid-large orgs (that are ahead of the AI curve) paints a different picture I honestly feel like predictions are made for the sake of seeming like some intelligent oracle. It’s easy to say things are > — [MrCollison on x · 1 points, 1 comments](https://x.com/MrCollison/status/2104984535888794050) **Source threads** - [x](https://x.com/housecor/status/2104969416039374964) · 0 points · 87 comments ## Similar posts on daily.dev - [System 1 vs. System 2 Agent Harnesses, clearly explained](https://daily.dev/posts/system-1-vs-system-2-agent-harnesses-clearly-explained-i8saxylfz) · Daily Dose of Data Science \| Avi Chawla \| Substack · 0 upvotes · 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#ai-agents](https://daily.dev/tags/ai-agents), [#prompt-engineering](https://daily.dev/tags/prompt-engineering), [#salesforce](https://daily.dev/tags/salesforce) [View this post on daily.dev](https://daily.dev/posts/tuning-agent-prompts-matters-more-than-model-choice-salesforce-research-finds-k50dlsbqn) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Tuning agent prompts matters more than model choice, Salesforce research finds","url":"https://daily.dev/posts/tuning-agent-prompts-matters-more-than-model-choice-salesforce-research-finds-k50dlsbqn","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/tuning-agent-prompts-matters-more-than-model-choice-salesforce-research-finds-k50dlsbqn"},"datePublished":"2026-09-22T18:24:55.458Z","dateModified":"2026-09-30T03:36:47.881Z","description":"A Salesforce AI research paper finds that once an agent's system prompt, tools, and workflow are carefully tuned around a weaker model, swapping in a stronger...","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/tuning-agent-prompts-matters-more-than-model-choice-salesforce-research-finds-k50dlsbqn","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,ai-agents,prompt-engineering,salesforce","timeRequired":"PT5M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Tuning agent prompts matters more than model choice, Salesforce research finds"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/tuning-agent-prompts-matters-more-than-model-choice-salesforce-research-finds-k50dlsbqn#faq","mainEntity":[{"@type":"Question","name":"Does upgrading to a stronger LLM automatically improve an AI agent's performance?","acceptedAnswer":{"@type":"Answer","text":"Not necessarily. Salesforce AI research found that once an agent's system prompt, tools, and workflow are carefully tuned around a weaker model, simply swapping in a stronger model does not automatically yield better results. The harness—prompt structure, tool definitions, and workflow logic—carries substantial weight in overall performance, sometimes more than the underlying model's raw capability. daily.dev surfaces research like this for engineers deciding whether to upgrade models or retune their agent harness."}},{"@type":"Question","name":"Why might a benchmark comparing different LLMs inside an agent loop be misleading?","acceptedAnswer":{"@type":"Answer","text":"Because if the prompt harness isn't separately tuned for each model, the benchmark ends up measuring prompt fit rather than true model capability. Salesforce research emphasizes that system prompts, tool definitions, and workflow logic are load-bearing components of agent performance, so an untuned harness can unfairly favor whichever model the prompt happened to be optimized for. Teams comparing agent models track findings like this on daily.dev before trusting a benchmark."}}]} ```