# How Smaller Tool Sets and Context Improved an Internal Incident Agent

DevFeed: [How Smaller Tool Sets and Context Improved an Internal Incident Agent](<https://devfeed.tech/articles/the-best-ai-fix-isn-t-adding-more-it-s-taking-away-64403.md>)

Original publisher: [Read original article](<https://hackernoon.com/the-best-ai-fix-isnt-adding-more-its-taking-away?source=rss>)

Author: hacker86245963

Published: 2026-10-03T02:11:06Z

Content type: article

Language: en

Sources: [HackerNoon](<https://devfeed.tech/sources/hackernoon.md>)

Topics: [AI Context Management](<https://devfeed.tech/topics/ai-context-management.md>), [context window](<https://devfeed.tech/topics/context-window.md>), [ReAct agent Python](<https://devfeed.tech/topics/react-agent-python.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-agent](<https://devfeed.tech/tags/ai-agent.md>), [ai-agent-architecture](<https://devfeed.tech/tags/ai-agent-architecture.md>), [ai-agents](<https://devfeed.tech/tags/ai-agents.md>), [ai-agents-guide](<https://devfeed.tech/tags/ai-agents-guide.md>), [ai-agents-tutorial](<https://devfeed.tech/tags/ai-agents-tutorial.md>), [ai-context-window](<https://devfeed.tech/tags/ai-context-window.md>), [ai-tips-and-tricks](<https://devfeed.tech/tags/ai-tips-and-tricks.md>), [context-window](<https://devfeed.tech/tags/context-window.md>), [debugging](<https://devfeed.tech/tags/debugging.md>), [engineering](<https://devfeed.tech/tags/engineering.md>), [eval](<https://devfeed.tech/tags/eval.md>), [fast](<https://devfeed.tech/tags/fast.md>), [filesystem](<https://devfeed.tech/tags/filesystem.md>), [latency](<https://devfeed.tech/tags/latency.md>), [llm](<https://devfeed.tech/tags/llm.md>), [logging](<https://devfeed.tech/tags/logging.md>), [team](<https://devfeed.tech/tags/team.md>)

## AI overview

An engineering team improved an internal AI agent for production incident triage by reducing its tools from 22 to 6, loading only relevant tools for each step, and keeping raw tool output out of the conversation context. They also compacted history and delegated parallel checks to subagents. On past incidents, average context per turn fell from about 75,000 tokens to 18,000, median response time dropped from 28 seconds to roughly 7, and the internal evaluation score rose from the low 60s to the mid 80s.

## Source excerpt

I noticed something interesting. Whenever an AI agent starts doing strange things, the immediate response of any team is to add more of something.