SurvivalIndex Measures Which Dev Tools AI Agents Actually Choose to Use
What Happened
A new project called SurvivalIndex is benchmarking something most tool makers haven't thought to measure: when AI coding agents are given a task with no tool name hints, what software do they actually reach for?
The methodology works like this: standardized repos are set up with natural-language prompts that describe what needs to be done without naming any specific tool. AI coding agents then solve the problem, and SurvivalIndex tracks which tools, libraries, and approaches they select. No hints, no steering - just "do this thing" and see what happens.
The early findings are notable. Claude Code chose custom/DIY implementations in 12 out of 20 tool categories tested. Not because it couldn't use existing tools - capability benchmarks (BFCL scores) suggest it can. It just doesn't reach for them. The project calls this a different failure mode than what capability benchmarks measure: it's not "can the agent use the tool" but "does the agent know the tool exists and choose it."
SurvivalIndex scores each tool across six dimensions: Insight Compression (embedded hard-won knowledge), Substrate Efficiency (computational resources), Broad Utility (cross-domain use), Awareness (documentation quality), Agent Friction (ease of AI integration), and Human Coefficient (embedded human judgment). Tools receive scores from 0-10, with tiers ranging from S-Tier "Immortal" (9.0+, e.g., Linux Kernel at 9.8, Git at 9.6) down to F-Tier "Endangered" (below 5.0).
Why It Matters
This research surfaces a blind spot in how we evaluate AI coding tools. Current benchmarks test whether an agent can use a tool when told to. SurvivalIndex tests whether the agent voluntarily selects that tool when solving a real problem. Those are fundamentally different questions.
The Claude Code finding is the most interesting data point. If the leading coding agent rebuilds from scratch in 60% of categories rather than using established tools, that has implications for the entire developer tools industry. Tools that AI agents don't spontaneously discover and use will see declining adoption as more development work flows through AI assistants.
This also matters for tool makers. If your product has poor "agent visibility" - meaning AI models don't know about it or don't think to use it - you have a distribution problem that no amount of marketing to humans will fix. The customer is increasingly an AI agent, not a developer browsing Product Hunt.
Our Take
SurvivalIndex is measuring something important, even if the dataset is still small. The 12-out-of-20 custom/DIY stat for Claude Code matches what we see in practice: coding agents tend to inline solutions rather than reaching for dependencies. Sometimes that's the right call - fewer dependencies means fewer supply chain risks. But often it means the agent is writing a worse version of something that already exists.
This has real consequences for tool selection. If you're evaluating developer tools today, you should be asking: "Will my AI coding agent actually use this?" A tool with perfect human usability but zero agent visibility has a shrinking addressable market.
The six-dimension scoring framework is thoughtful, particularly the "Agent Friction" metric. Tools that ship with machine-readable documentation, clear API schemas, and minimal setup steps will have a structural advantage as AI-mediated development grows. If you maintain an open-source project, your README and docs aren't just for human developers anymore - they're retrieval targets for coding agents.
It's early research with a small sample size, so don't restructure your toolchain based on it yet. But the gap between "tools agents can use" and "tools agents choose to use" is going to matter a lot more in the next year.