--- title: Evolution of agentic surfaces — Gagan Bhat & Isabella... description: Anthropic engineers Gagan Bhat and Isabella Kai He walk through three years of evolution in agentic surfaces: from the Messages API (tokens in/out), to the... canonical: https://daily.dev/posts/evolution-of-agentic-surfaces-gagan-bhat-isabella-kai-he-anthropic-5ilu2afcz twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Evolution of agentic surfaces — Gagan Bhat & Isabella Kai He, Anthropic | daily.dev og:description: Anthropic engineers Gagan Bhat and Isabella Kai He walk through three years of evolution in agentic surfaces: from the Messages API (tokens in/out), to the... og:url: https://daily.dev/posts/evolution-of-agentic-surfaces-gagan-bhat-isabella-kai-he-anthropic-5ilu2afcz og:image: https://api.daily.dev/og/posts/5ILu2Afcz.png og:image:alt: Evolution of agentic surfaces — Gagan Bhat & Isabella Kai He, Anthropic 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. # Evolution of agentic surfaces — Gagan Bhat & Isabella Kai He, Anthropic **[AI Engineer](https://daily.dev/sources/aidotengineer)** · 31 min read · 0 upvotes · 0 comments ## Summary Anthropic engineers Gagan Bhat and Isabella Kai He walk through three years of evolution in agentic surfaces: from the Messages API (tokens in/out), to the Claude Agent SDK (built-in agentic loop), to Claude Managed Agents — a fully managed production infrastructure layer. Claude Managed Agents decouples the 'brain' (agent loop) from the 'hands' (sandbox/tool execution), yielding 60% faster time-to-first-token at P50 and over 90% improvement at P95. The talk covers core primitives (agent, environment, session), engineering principles like harness agility as models evolve, and production lessons: credential vaults, session logs for observability and self-improvement, self-hosted sandboxes for VPC-bound enterprises, and MCP tunnels for private network MCP servers. Frontier features discussed include 'dreaming' (periodic batch processing of session logs to update agent memory and enable self-improvement) and 'outcomes' (a grader agent that checks success criteria and retries until the task is complete). ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What is Claude Managed Agents and how does it differ from the Claude Agent SDK? Claude Managed Agents is Anthropic's fully managed agent infrastructure that decouples the agent's reasoning loop (the brain) from its tool execution sandbox (the hands), with Anthropic running hosting, session management, credentials, and observability. This differs from the Claude Agent SDK, which packages a built-in agentic loop and Claude Code's harness but still requires developers to hand-roll credentials, hosting infrastructure, and scaling themselves. _Engineers evaluating agent infrastructure options can track architecture shifts like this on daily.dev._ ### How much faster is time to first token after decoupling the agent brain from the sandbox execution environment? Decoupling the reasoning loop from the tool execution sandbox produced 60% faster time to first token for median (P50) use cases and over 90% improvement for P95 use cases, because the model can begin reasoning immediately instead of waiting for container setup to finish, running setup in parallel or skipping it entirely when unneeded. _Teams tuning agent latency can follow architectural benchmarks like this via daily.dev._ ### What was the context anxiety issue in Claude Sonnet 4.5 and how was it fixed? Context anxiety was a behavior where Sonnet 4.5 wrapped up tasks and terminated work early as it approached its context window limit, even with room to spare. Anthropic's team added context reset fixes into the agent harness to compensate, but when Opus 4.5 shipped, the behavior disappeared entirely, turning those same fixes into pure overhead that added latency and caused cache invalidation issues. _Developers maintaining custom agent harnesses can watch for model-specific quirks like this on daily.dev._ ## Similar posts on daily.dev - [Anthropic Introduces Managed Agents to Simplify AI Agent Deployment](https://daily.dev/posts/anthropic-introduces-managed-agents-to-simplify-ai-agent-deployment-em1vdc2kw) · InfoQ · 1 upvotes · 0 comments - [Introducing Claude Managed Agents with Modal Sandboxes](https://daily.dev/posts/introducing-claude-managed-agents-with-modal-sandboxes-oqf9ocaau) · Modal · 0 upvotes · 0 comments --- Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#claude](https://daily.dev/tags/claude), [#anthropic](https://daily.dev/tags/anthropic), [#context-engineering](https://daily.dev/tags/context-engineering) [View this post on daily.dev](https://daily.dev/posts/evolution-of-agentic-surfaces-gagan-bhat-isabella-kai-he-anthropic-5ilu2afcz) ```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":"Evolution of agentic surfaces — Gagan Bhat & Isabella Kai He, Anthropic","url":"https://daily.dev/posts/evolution-of-agentic-surfaces-gagan-bhat-isabella-kai-he-anthropic-5ilu2afcz","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/evolution-of-agentic-surfaces-gagan-bhat-isabella-kai-he-anthropic-5ilu2afcz"},"datePublished":"2026-08-11T16:19:43.531Z","dateModified":"2026-09-14T07:30:32.341Z","description":"Anthropic engineers Gagan Bhat and Isabella Kai He walk through three years of evolution in agentic surfaces: from the Messages API (tokens in/out), to the...","image":"https://i.ytimg.com/vi/K0X9QDRkIdg/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/K0X9QDRkIdg/sddefault.jpg","isAccessibleForFree":true,"articleSection":"AI Engineer","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":"AI Engineer","logo":"https://media.daily.dev/image/upload/s--u5PucxNT--/f_auto/v1724338940/logos/aidotengineer","url":"https://daily.dev/sources/aidotengineer"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/evolution-of-agentic-surfaces-gagan-bhat-isabella-kai-he-anthropic-5ilu2afcz","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai-agents,claude,anthropic,context-engineering","timeRequired":"PT31M","video":{"@type":"VideoObject","name":"Evolution of agentic surfaces — Gagan Bhat & Isabella Kai He, Anthropic","description":"Anthropic engineers Gagan Bhat and Isabella Kai He walk through three years of evolution in agentic surfaces: from the Messages API (tokens in/out), to the...","thumbnailUrl":"https://i.ytimg.com/vi/K0X9QDRkIdg/sddefault.jpg","uploadDate":"2026-08-11T16:19:43.531Z","duration":"PT31M","url":"https://api.daily.dev/r/5ILu2Afcz","embedUrl":"https://www.youtube.com/embed/K0X9QDRkIdg"}} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"AI Engineer","item":"https://daily.dev/sources/aidotengineer"},{"@type":"ListItem","position":3,"name":"Evolution of agentic surfaces — Gagan Bhat & Isabella Kai He, Anthropic"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/evolution-of-agentic-surfaces-gagan-bhat-isabella-kai-he-anthropic-5ilu2afcz#faq","mainEntity":[{"@type":"Question","name":"What is Claude Managed Agents and how does it differ from the Claude Agent SDK?","acceptedAnswer":{"@type":"Answer","text":"Claude Managed Agents is Anthropic's fully managed agent infrastructure that decouples the agent's reasoning loop (the brain) from its tool execution sandbox (the hands), with Anthropic running hosting, session management, credentials, and observability. This differs from the Claude Agent SDK, which packages a built-in agentic loop and Claude Code's harness but still requires developers to hand-roll credentials, hosting infrastructure, and scaling themselves. Engineers evaluating agent infrastructure options can track architecture shifts like this on daily.dev."}},{"@type":"Question","name":"How much faster is time to first token after decoupling the agent brain from the sandbox execution environment?","acceptedAnswer":{"@type":"Answer","text":"Decoupling the reasoning loop from the tool execution sandbox produced 60% faster time to first token for median (P50) use cases and over 90% improvement for P95 use cases, because the model can begin reasoning immediately instead of waiting for container setup to finish, running setup in parallel or skipping it entirely when unneeded. Teams tuning agent latency can follow architectural benchmarks like this via daily.dev."}},{"@type":"Question","name":"What was the context anxiety issue in Claude Sonnet 4.5 and how was it fixed?","acceptedAnswer":{"@type":"Answer","text":"Context anxiety was a behavior where Sonnet 4.5 wrapped up tasks and terminated work early as it approached its context window limit, even with room to spare. Anthropic's team added context reset fixes into the agent harness to compensate, but when Opus 4.5 shipped, the behavior disappeared entirely, turning those same fixes into pure overhead that added latency and caused cache invalidation issues. Developers maintaining custom agent harnesses can watch for model-specific quirks like this on daily.dev."}}]} ```