Krishan's Tech Blog
Empirical 642-PR Dataset & Real-Time Classifier

AI Agent Bloat & Deletion Report Card

Autonomous software agents (Claude Code, Devin, Copilot SWE) solve prompts by adding code rather than deleting or refactoring. Pick any studied repo or enter any public GitHub repository to grade its AI deletion deficit.

Preloaded Study Repositories (Click to inspect)

BerriAI/litellm PRELOADED STUDY

Empirical benchmark of 40 agent PRs (Devin) matched 1:1 against 40 human PRs.

C Bloat Grade

Grade breakdown by task type

Type score = geometric mean of three agent-vs-human ratios (net lines, deletions with human over agent, net lines per file). Each ratio is clamped to 0.5-2 (▲ marks a clamped one). Above 1 is worse for the agent.

Task mix (% of each cohort's PRs)

Each PR is tagged feature, fix, test, docs, refactor, chore or other from its title prefix, labels or wording. The grade compares agents with humans within each type that has at least 3 agent and 3 human PRs (marked ✓): median net lines, median deletions and net lines per file, weighted by agent PRs. Each of the three ratios is clamped to between 0.5 and 2 first. A is a score of 0.8 or lower, B up to 1.1x, C up to 1.5x, D above. It does not count how often a PR grows code, because feature PRs normally do.

Net-Growth PR % (context)
🤖 Agent: 75.0%
👤 Human: 62.5%
Zero-Deletion PR %
🤖 Agent: 20.0%
👤 Human: 10.0%
Median Deletions / PR
🤖 Agent: 6.5 lines
👤 Human: 17.5 lines
Median Additions / PR
🤖 Agent: 142 lines
👤 Human: 110 lines
Median Files Changed / PR
🤖 Agent: -
👤 Human: -
Median Commits / PR
🤖 Agent: -
👤 Human: -
PRs With Review Comments
🤖 Agent: -
👤 Human: -
Median Time to Merge
🤖 Agent: -
👤 Human: -
Fix agent bloat before it merges with Ratchet

Enforce code size caps directly in CI so AI agents are forced to refactor & delete dead code.

pip install ratchet-size && ratchet init
View on GitHub

Sample Analyzed PRs

Cohort PR Title Author / Agent Changes