๐ญ Something interesting happened while I was building a long-running AI agent workflow today.
Like most people, I always assumed that if an AI agent starts slowing down or behaving strangely, it's probably because the context window is getting full.
Turns out, it's not that simple.
๐ณ๏ธ My workflow started freezing even though the context utilization was only around ๐๐โ๐๐%. That sent me down a rabbit hole to understand what was actually happening under the hood.
The more I read, the more I realized that context size is just one piece of the puzzle. Long-running agent sessions can also be affected by:
โ๏ธ ๐๐จ๐จ๐ฅ-๐๐๐ฅ๐ฅ ๐จ๐ฏ๐๐ซ๐ก๐๐๐ โ Every tool invocation introduces another request-response cycle, and dozens of chained calls can add noticeable latency.
๐ฆ ๐๐๐ซ๐ ๐ ๐ข๐ง๐ญ๐๐ซ๐ฆ๐๐๐ข๐๐ญ๐ ๐ญ๐จ๐จ๐ฅ ๐จ๐ฎ๐ญ๐ฉ๐ฎ๐ญ๐ฌ โ Reading many files, search results, or diffs can flood the agent with information long before the context window is full.
๐ง ๐๐จ๐ง๐ญ๐๐ฑ๐ญ ๐ซ๐๐๐จ๐ง๐ฌ๐ญ๐ซ๐ฎ๐๐ญ๐ข๐จ๐ง โ After each tool call, the model often has to rebuild conversational state, which becomes increasingly expensive as the session grows.
๐ ๐๐ ๐๐ง๐ญ ๐จ๐ซ๐๐ก๐๐ฌ๐ญ๐ซ๐๐ญ๐ข๐จ๐ง ๐ฅ๐ข๐ฆ๐ข๐ญ๐๐ญ๐ข๐จ๐ง๐ฌ โ Timeouts, retries, scheduling, and orchestration logic can become bottlenecks in complex workflows.
โก ๐๐ ๐๐๐๐ก๐ ๐ข๐ง๐ฏ๐๐ฅ๐ข๐๐๐ญ๐ข๐จ๐ง/๐๐ฏ๐ข๐๐ญ๐ข๐จ๐ง โ Losing the cached attention state can increase recomputation, making subsequent interactions slower and less responsive.
๐คฏ It definitely changed the way I'm thinking about building AI agents. There's a lot more happening under the hood than just the context window.
If you're building with @cursor_ai , @claudeai Code, Codex, or any other agentic coding tool, I'd genuinely love to know if you've run into something similar or discovered other bottlenecks that aren't immediately obvious.
๐ I'm also putting together a more detailed blog where I'll dive deeper into why this happens and, more importantly, the architectural approaches and practical solutions that helped me make my workflows more stable and responsive.
#AI #GenAI #AIAgents #AgenticAI #LLM #HarnessEngineering #Cursor #ClaudeCode #Codex #MCP #AgentEngineering #ContextEngineering #LLMOps #SoftwareEngineering #BuildInPublic
7/ This feels like MCP-style tool exposure,
That changes app discovery.
The question may shift from:
โCan users find my app?โ to: โCan agents safely use my app?โ
It feels like Android is becoming an intelligence layer for apps, agents, UI, and devices. ๐
Google I/O 2026 felt less like a product eventโฆ and more like the start of a new Android era. ๐
Apps are moving from: ๐ฑ app-first experiences to ๐ค intent-first experiences
As an Android engineer, this one felt serious. ๐งต
#AndroidDev#GoogleIO#Gemini#AgenticAI#AI
@phonescloud99 Another issue to map your project's actual files with its correct figma frame.. this doesn't workout at all.. and you have to take care of tha manually
We donโt have a coding problem anymore.
We have a harness problem.
That thought pushed me to build Screen Agent, an Android + AI agent workflow that starts from the real device, captures runtime context, and closes the loop back to code.
I wrote about the journey here ๐
https://t.co/u0PpVPapnoโ
#Android #AgenticAI #AIEngineering #DeveloperTools #HarnessEnginerring
@phonescloud99 Agree.. I am still figuring out the right approach to match the Figma's accuracy, as the tools provided by Figma MCP are unable to match tge expectations.
It can provide you the Metadata of any frame, but the final alignment is not always satisfactory.