[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"nav-notes-count":3,"content-note-when-should-a-coding-agent-use-a-more-capable-model":4},9,{"id":5,"title":6,"author":7,"authorAvatar":8,"authorRole":9,"body":10,"category":311,"date":312,"description":94,"excerpt":313,"extension":314,"image":315,"index":316,"meta":317,"navigation":330,"path":331,"pullQuote":332,"readingTime":333,"seo":334,"slug":335,"stem":336,"takeaways":337,"total":341,"__hash__":342},"notes\u002Fnotes\u002Fwhen-should-a-coding-agent-use-a-more-capable-model.md","When Should a Coding Agent Use a More Capable Model?","Norbert Brett","\u002Fimages\u002Fnorbert-brett.jpg","Full-Stack & AI Systems Engineer",{"type":11,"value":12,"toc":307},"minimark",[13,18,22,70,81,88,123,132,216,220,228,245,248,292,303],[14,15,17],"h2",{"id":16},"scoring-task-risk-before-burning-frontier-tokens","Scoring Task Risk Before Burning Frontier Tokens",[19,20,21],"p",{},"Every few weeks, a new AI model launch trends on X and half of the tech timeline declares that model routing is obsolete because the newest flagship model can do everything. Then you actually run that flagship model across a 60-file software project, watch it spend 90 seconds \"thinking\" just to add a button style, and hit your usage quota before lunch. Months earlier, I was frustrated by that exact trap: I was burning my most expensive, highest-latency reasoning model on repetitive formatting tasks, leaving zero capacity for the database and security decisions that actually required deep thought.",[23,24,25,32],"blockquote",{},[19,26,27,31],{},[28,29,30],"strong",{},"In Plain English — What is \"Selective Model Routing\"?","\nAI models come in different sizes and speeds, much like vehicles in a logistics fleet:",[33,34,35,47,56],"ul",{},[36,37,38,46],"li",{},[28,39,40,41,45],{},"Frontier Reasoning Models (e.g., ",[42,43,44],"code",{},"Gemini Pro \u002F Argon",")",": Like a senior structural engineer—capable of weighing complex trade-offs across an entire building blueprint, but slower and 10–30x more expensive per page.",[36,48,49,55],{},[28,50,51,52,45],{},"Fast Execution Models (e.g., ",[42,53,54],{},"Gemini Flash \u002F Flash-Lite",": Like a skilled, high-speed installation crew—once you hand them an exact blueprint and measurements, they execute the work in seconds at a fraction of the cost.",[36,57,58,61,62,65,66,69],{},[28,59,60],{},"Selective Model Routing",": Automatically choosing which model handles each step based on ",[28,63,64],{},"how ambiguous"," and ",[28,67,68],{},"how risky"," the task is.",[19,71,72,73,76,77,80],{},"When I built the dynamic routing engine in ",[28,74,75],{},"Gemini Advisor"," (",[42,78,79],{},"scripts\u002Fmodel_allocator.sh","), I made risk scoring mandatory before any file is edited:",[19,82,83],{},[84,85],"img",{"alt":86,"src":87},"Selective Model Routing Architecture","\u002Fimages\u002Fnotes\u002Fdiagrams\u002Fwhen-should-a-coding-agent-use-a-more-capable-model.svg",[89,90,95],"pre",{"className":91,"code":92,"language":93,"meta":94,"style":94},"language-bash shiki shiki-themes github-light github-dark","# Dynamic tier check before modifying workspace files\n.\u002Fscripts\u002Fmodel_allocator.sh --check\n# Emits: SELECTIVE ROUTE mode=\"solo|delegate|audit|full\" risk=\"low|medium|high\"\n","bash","",[42,96,97,106,117],{"__ignoreMap":94},[98,99,102],"span",{"class":100,"line":101},"line",1,[98,103,105],{"class":104},"sJ8bj","# Dynamic tier check before modifying workspace files\n",[98,107,109,113],{"class":100,"line":108},2,[98,110,112],{"class":111},"sScJk",".\u002Fscripts\u002Fmodel_allocator.sh",[98,114,116],{"class":115},"sj4cs"," --check\n",[98,118,120],{"class":100,"line":119},3,[98,121,122],{"class":104},"# Emits: SELECTIVE ROUTE mode=\"solo|delegate|audit|full\" risk=\"low|medium|high\"\n",[19,124,125,126,65,129],{},"You see, the dividing line between tasks that require a Frontier reasoning model and tasks that run better on a fast execution model comes down to two questions: ",[28,127,128],{},"Is the design still ambiguous?",[28,130,131],{},"What is the blast radius if a mistake happens?",[133,134,135,160],"table",{},[136,137,138],"thead",{},[139,140,141,146,153],"tr",{},[142,143,145],"th",{"align":144},"left","Decision Factor",[142,147,148,149,152],{"align":144},"Use Frontier Tier (",[42,150,151],{},"Pro"," \u002F Architect)",[142,154,155,156,159],{"align":144},"Use Execution Tier (",[42,157,158],{},"Flash"," \u002F Builder)",[161,162,163,177,190,203],"tbody",{},[139,164,165,171,174],{},[166,167,168],"td",{"align":144},[28,169,170],{},"Clarity of Spec",[166,172,173],{"align":144},"Requirements are open-ended; several different designs could work.",[166,175,176],{"align":144},"File paths, data types, and rules are already locked in a written spec.",[139,178,179,184,187],{},[166,180,181],{"align":144},[28,182,183],{},"Blast Radius (Risk)",[166,185,186],{"align":144},"Touches login security, payment logic, database schemas, or shared routing.",[166,188,189],{"align":144},"Confined to 1–3 specific UI components, test files, or content pages.",[139,191,192,197,200],{},[166,193,194],{"align":144},[28,195,196],{},"Cross-File Context",[166,198,199],{"align":144},"Must connect dots across 15+ files or audit a finished diff for subtle bugs.",[166,201,202],{"align":144},"Only needs to read the target file and the 5-part implementation contract.",[139,204,205,210,213],{},[166,206,207],{"align":144},[28,208,209],{},"Cost of an Error",[166,211,212],{"align":144},"Silent data leak, broken customer checkout, or security hole.",[166,214,215],{"align":144},"Immediate compiler or test error that gets caught and fixed in 3 seconds.",[14,217,219],{"id":218},"measuring-the-trade-off-why-fast-models-need-a-5-part-contract","Measuring the Trade-Off: Why Fast Models Need a 5-Part Contract",[19,221,222,223,227],{},"If you give a fast execution model a vague five-word prompt like ",[224,225,226],"em",{},"\"Add rate limiting to the contact form,\""," it will guess—and you will spend an hour cleaning up its assumptions. If you have ever blamed a faster model for \"bad code\" after giving it a one-sentence instruction, the model was not the problem; the missing blueprint was.",[19,229,230,231,233,234,237,238,241,242,244],{},"In my workflow across ",[28,232,75],{}," and the ",[28,235,236],{},"Book-to-Skill"," distiller, the Frontier model never disappears—it just stops typing out 400 lines of repetitive implementation code. Instead, it writes the ",[28,239,240],{},"5-Part Contract"," (exact data shapes, allowed files, forbidden shortcuts, and test commands) and hands that blueprint to a ",[42,243,158],{}," builder subagent.",[19,246,247],{},"When we measured this split across 20 real feature and refactoring tasks in 2026, three numbers proved the trade-off:",[249,250,251,264,283],"ol",{},[36,252,253,256,257,259,260,263],{},[28,254,255],{},"15x Reduction in Frontier Token Usage",": Because trial-and-error compiler fixes happen inside the cheap, disposable ",[42,258,158],{}," builder session instead of cluttering the main architect's memory, parent context usage stayed below ",[28,261,262],{},"25%",".",[36,265,266,269,270,272,273,276,277,279,280,263],{},[28,267,268],{},"4-Minute Parallel Processing",": In ",[28,271,236],{},", after a deterministic Python script (",[42,274,275],{},"extract.py",") calculates exact chapter byte offsets, parallel waves of ",[42,278,158],{}," subagents distill a 500-page technical manual into structured agent skills in under ",[28,281,282],{},"4 minutes",[36,284,285,288,289,291],{},[28,286,287],{},"Automatic Safety Escalation",": If a ",[42,290,158],{}," builder fails the automated test suite twice or hits an unexpected schema conflict, the router automatically escalates back to the Frontier model rather than letting the fast model guess.",[19,293,294,295,298,299,302],{},"I felt relieved once I stopped asking ",[224,296,297],{},"\"Which single AI model is best on social media?\""," and started measuring ",[28,300,301],{},"verified cost per merged feature",". Use your deepest reasoning model to lock the blueprint and inspect the result, and let the fast model do the heavy typing in between.",[304,305,306],"style",{},"html pre.shiki code .sJ8bj, html code.shiki .sJ8bj{--shiki-default:#6A737D;--shiki-dark:#6A737D}html pre.shiki code .sScJk, html code.shiki .sScJk{--shiki-default:#6F42C1;--shiki-dark:#B392F0}html pre.shiki code .sj4cs, html code.shiki .sj4cs{--shiki-default:#005CC5;--shiki-dark:#79B8FF}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}",{"title":94,"searchDepth":108,"depth":108,"links":308},[309,310],{"id":16,"depth":108,"text":17},{"id":218,"depth":108,"text":219},"Engineering Craft","November 1, 2026",null,"md","\u002Fimages\u002Fnotes\u002Fwhen-should-a-coding-agent-use-a-more-capable-model.webp","04",{"sections":318},[319,324],{"marker":320,"heading":17,"body":321},"\u002F A",[322,323],"Every few weeks a new model release dominates X with claims that it makes every previous workflow obsolete. Yet if you run autonomous coding agents daily, you hit a more practical problem: burning your deepest reasoning tier on mechanical boilerplate exhausts your rate limits in two hours.","In Gemini Advisor, model_allocator.sh inspects active tiers dynamically and enforces an upfront SELECTIVE ROUTE declaration—routing low-risk bounded edits to Gemini Flash while reserving Frontier capacity for architecture and zero-history diff audits.",{"marker":325,"heading":326,"body":327},"\u002F B","Measuring the 15x Cost-to-Regression Trade-Off",[328,329],"A fast model fails when given a vague prompt, not when given a locked type signature. When the Frontier planner supplies a 5-part contract with explicit file boundaries, Flash completes mechanical implementation slices at 10–30x lower token cost.","The metric that matters is not raw tokens per second; it is verified cost per merged pull request with zero post-merge reverts.",true,"\u002Fnotes\u002Fwhen-should-a-coding-agent-use-a-more-capable-model","Using a frontier reasoning model to rename forty CSS variables is like hiring a structural architect to carry drywall up the stairs.","7 min read",{"title":6,"description":94},"when-should-a-coding-agent-use-a-more-capable-model","notes\u002Fwhen-should-a-coding-agent-use-a-more-capable-model",[338,339,340],"Match the AI model to the risk of the task: use deep reasoning models for architecture and security, and fast models for bounded code edits.","Fast models rarely fail because they are 'dumb'—they fail when given vague instructions instead of a written 5-part contract.","Across 20 real engineering tasks, selective routing cut frontier token costs by 15x while keeping parent context usage under 25%.","10","0nixxZ0GKXTJGGCqrpoiBKpBAgu-B2cUOFsstgB_6zE"]