--- title: Talk: Type Checking in Agentic Workflows | daily.dev description: Pyrefly's team presented experimental findings at the PyCon US 2026 Typing Conference on whether adding a type checker to agentic coding workflows improves... canonical: https://daily.dev/posts/talk-type-checking-in-agentic-workflows-c6uywhawz twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Talk: Type Checking in Agentic Workflows | daily.dev og:description: Pyrefly's team presented experimental findings at the PyCon US 2026 Typing Conference on whether adding a type checker to agentic coding workflows improves... og:url: https://daily.dev/posts/talk-type-checking-in-agentic-workflows-c6uywhawz og:image: https://api.daily.dev/og/posts/c6uyWHAWZ.png og:image:alt: Talk: Type Checking in Agentic Workflows 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. # Talk: Type Checking in Agentic Workflows **[Pyrefly: A Fast Python Type Checker and Language Server \| Pyrefly](https://daily.dev/sources/pyrefly)** · 8 min read · 0 upvotes · 0 comments ## Summary Pyrefly's team presented experimental findings at the PyCon US 2026 Typing Conference on whether adding a type checker to agentic coding workflows improves task success. Type checking helped only when the codebase was already well-typed, raising success rate from about 80% to 84% and reducing steps and duration; in low-coverage codebases, type errors distracted agents into chasing checker cleanliness instead of the actual task. How feedback is delivered mattered as much as the feedback itself: per-edit checks surfaced as a separate conversation step worked best, models differed in sensitivity (Claude reactive, GPT Codex more goal-oriented needing structure), and frequent, consistent feedback prevented unproductive 'search spirals'. Experiments used SWE-bench Verified (low type coverage, Claude Sonnet 4.5, GPT-5.3) and an internal MetaSWEBench with Pyre (higher type coverage, Claude Opus). Open questions include testing other type checkers, well-typed corpora, frontier models, and more type-relevant task types. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### Does adding a type checker actually improve AI coding agent success rates? It depends on how well-typed the codebase already is. In well-typed codebases, adding a type checker to an agentic workflow raised task success rate from about 80% to 84%, reduced the number of steps, and shortened task duration. In codebases with low type coverage, such as SWE-bench Verified's Matplotlib, Django, and SymPy tasks, type errors distracted agents into fixing unrelated import or signature issues instead of the actual task, with no meaningful improvement. _daily.dev surfaces findings like this for teams deciding whether type checking pays off in agentic coding pipelines._ ### How should type checker feedback be delivered to an AI coding agent for best results? Feedback works best delivered as a separate conversation step after each edit, rather than bundled into the edit success message, and it should run consistently on every edit rather than only once at the end. High-frequency, consistent feedback prevents unproductive 'search spirals' where the model repeatedly re-verifies prior work, and even a clean 'zero errors found' message helps keep agents on track. _Engineers tuning agentic coding loops track practical feedback-design lessons like this on daily.dev._ ### Do different AI models respond differently to type checker error feedback during coding tasks? Yes, Claude Sonnet 4.5 and Claude Opus were highly sensitive to type errors, fixing whatever was shown but also getting derailed by noisy signals, while GPT-5.3 (Codex) was more goal-oriented and required structural intervention to actually address surfaced errors. These sensitivities are expected to shift as models progress, affecting how aggressively errors should be filtered. _daily.dev helps developers comparing model behavior stay current as agentic coding models evolve._ ## Similar posts on daily.dev - [Catching up with the Python Typing Council](https://daily.dev/posts/catching-up-with-the-python-typing-council-a4c2xw3fs) · Planet Python · 2 upvotes · 0 comments - [Pyrefly Now Type Checks PyTorch – PyTorch](https://daily.dev/posts/pyrefly-now-type-checks-pytorch-pytorch-adohlgdfc) · PyTorch · 0 upvotes · 0 comments - [Python Type Checker Comparison: Typing Spec Conformance](https://daily.dev/posts/python-type-checker-comparison-typing-spec-conformance-yvjkxqkwd) · Hacker News · 1 upvotes · 0 comments - [10 tips to improve your coding agent game](https://daily.dev/posts/10-tips-to-improve-your-coding-agent-game-r2zthx2x0) · Pulumi · 3 upvotes · 0 comments --- Tags: [#ai](https://daily.dev/tags/ai), [#python](https://daily.dev/tags/python), [#ai-agents](https://daily.dev/tags/ai-agents) [View this post on daily.dev](https://daily.dev/posts/talk-type-checking-in-agentic-workflows-c6uywhawz) ```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":"Talk: Type Checking in Agentic Workflows","url":"https://daily.dev/posts/talk-type-checking-in-agentic-workflows-c6uywhawz","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/talk-type-checking-in-agentic-workflows-c6uywhawz"},"datePublished":"2026-08-23T12:17:28.036Z","dateModified":"2026-08-23T12:24:27.854Z","description":"Pyrefly's team presented experimental findings at the PyCon US 2026 Typing Conference on whether adding a type checker to agentic coding workflows improves...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/ce7cceb523e1bdb0e6d2d6310d2d101d?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/ce7cceb523e1bdb0e6d2d6310d2d101d?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Pyrefly: A Fast Python Type Checker and Language Server | Pyrefly","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":"Pyrefly: A Fast Python Type Checker and Language Server | Pyrefly","logo":"https://media.daily.dev/image/upload/s--Qzw5NpnF--/c_limit,w_256/f_auto,q_auto/v1787487420/logos/pyrefly?_a=BAMAMicg0","url":"https://daily.dev/sources/pyrefly"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/talk-type-checking-in-agentic-workflows-c6uywhawz","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai,python,ai-agents","timeRequired":"PT8M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Pyrefly: A Fast Python Type Checker and Language Server | Pyrefly","item":"https://daily.dev/sources/pyrefly"},{"@type":"ListItem","position":3,"name":"Talk: Type Checking in Agentic Workflows"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/talk-type-checking-in-agentic-workflows-c6uywhawz#faq","mainEntity":[{"@type":"Question","name":"Does adding a type checker actually improve AI coding agent success rates?","acceptedAnswer":{"@type":"Answer","text":"It depends on how well-typed the codebase already is. In well-typed codebases, adding a type checker to an agentic workflow raised task success rate from about 80% to 84%, reduced the number of steps, and shortened task duration. In codebases with low type coverage, such as SWE-bench Verified's Matplotlib, Django, and SymPy tasks, type errors distracted agents into fixing unrelated import or signature issues instead of the actual task, with no meaningful improvement. daily.dev surfaces findings like this for teams deciding whether type checking pays off in agentic coding pipelines."}},{"@type":"Question","name":"How should type checker feedback be delivered to an AI coding agent for best results?","acceptedAnswer":{"@type":"Answer","text":"Feedback works best delivered as a separate conversation step after each edit, rather than bundled into the edit success message, and it should run consistently on every edit rather than only once at the end. High-frequency, consistent feedback prevents unproductive 'search spirals' where the model repeatedly re-verifies prior work, and even a clean 'zero errors found' message helps keep agents on track. Engineers tuning agentic coding loops track practical feedback-design lessons like this on daily.dev."}},{"@type":"Question","name":"Do different AI models respond differently to type checker error feedback during coding tasks?","acceptedAnswer":{"@type":"Answer","text":"Yes, Claude Sonnet 4.5 and Claude Opus were highly sensitive to type errors, fixing whatever was shown but also getting derailed by noisy signals, while GPT-5.3 (Codex) was more goal-oriented and required structural intervention to actually address surfaced errors. These sensitivities are expected to shift as models progress, affecting how aggressively errors should be filtered. daily.dev helps developers comparing model behavior stay current as agentic coding models evolve."}}]} ```