Cole Halton

Cole Halton

Sep 20

The Gemini breakout verdict has to come from the boundary, not the model's mouth

#ai #security #agents #llm

5 min read

Cole Halton

Sep 20

The Gemini breakout is a judge problem, not a jailbreak problem

#ai #security #agents #eval

5 min read

Cole Halton

Sep 19

AGENTS.md is becoming the portability layer for coding agents, and it breaks reproducible compare

#codingagents #claudecode #benchmarking #codereview

5 min read

Cole Halton

Sep 19

An LLM reviewer's "block" is a feature, not a verdict

#aicodereview #llm #evaluation #machinelearning

1

4 min read

Cole Halton

Sep 18

You can read Claude Code's whole harness now. That's what every benchmark score throws away

#codingagents #harness #benchmark #eval

5 min read

Cole Halton

Sep 18

Verus proves Rust correct for all inputs. Code review still can't define "correct."

#aicodereview #rust #formalverification #llm

2 min read

Cole Halton

Sep 17

A student eyeballed 102 F-Droid apps for LLM slop. The method is the story.

#ai #llm #opensource

2 min read

Cole Halton

Sep 17

Reduce PR review time with AI: the bottleneck is the wait, not the read

#codereview #ai #productivity #engineering

1

2 min read

Cole Halton

Sep 16

Deleting a secret from your Docker image doesn't delete it from your build history

#cybersecurity #docker #github #security

2 min read

Cole Halton

Sep 16

A Docker container is containment, not a credential boundary

#aiagents #sandboxing #security #codereview

2 min read

Cole Halton

Sep 15

Cheap LLM code review is fine until it hits an authorization bug

#aicodereview #llm #security #evaluation

2 min read

Cole Halton

Sep 15

Why you shouldn't let the model review its own AI code

#aicodereview #agenticai #codereview #softwareengineering

2 min read

Cole Halton

Sep 14

Two "Codex CLI" models on the same benchmark: the harness hides the model

#ai #codingagents #benchmark #eval

2 min read

Cole Halton

Sep 14

Giving a coding agent more time barely helps

#agenticcoding #benchmarks #evaldesign #ai

1

2 min read

Cole Halton

Sep 13

The render said clean, the file said broken: why AI visual QA lies

#aireview #evaldesign #benchmarks #independentverification

2 min read

Cole Halton

Sep 13

The RubyGems agent attack is a coding-agent benchmark nobody writes

#agents #ai #cybersecurity #security

2

2 min read

Cole Halton

Sep 12

Stop asking the model that wrote the code to review it

#aicodereview #evaluation #aiagents #codereview

3

2 min read

Cole Halton

Sep 12

SWE-2's easy slice is near-perfect. Its scorecard hides why

#agents #ai #llm #machinelearning

1

2 min read

Cole Halton

Sep 11

Looped reasoning means the AI's visible trace isn't the reasoning

#agents #ai #llm #machinelearning

3

2 min read

Cole Halton

Sep 11

When the code you're reviewing isn't what the model wrote

#aicodereview #security #evaldesign

1

2 min read

Cole Halton

Sep 10

Microsoft's PRAssistant number is real, and it's a floor, not a promise

#aicodereview #pullrequests #ai

2 min read

Cole Halton

Sep 10

Atlassian says Rovo cut PR review time 45%. Here's the measurement they didn't publish.

#aicodereview #prreview #measurement #engineeringmanagement

2 min read

Cole Halton

Sep 9

The state axis: why agent benchmarks keep measuring amnesiac models

#agents #ai #llm

1

4

2 min read

Cole Halton

Sep 9

How to Actually Evaluate an AI Code Review Tool

#ai #codereview #engineering

1

2 min read

Cole Halton

Sep 8

Your team's coding rules aren't in the prompt, they're in the ingest

#ai #codereview #devtools

1

2 min read

Cole Halton

Sep 8

The GNU strip backdoor is the case AI code review can't see

#aicodereview #supplychain #security

2 min read

Cole Halton

Sep 7

The sleep loop is the tell: agents that pay per action optimize to do nothing

#aiagents #evaluation #llm #benchmarking

1

2 min read

Cole Halton

Sep 7

The model did the reverse-engineering. The validator was the hard part.

#aicoding #agents #llm #evaluation

1

2 min read

Cole Halton

Aug 20

How I actually eval AI code review tools (no vendor numbers)

#aicodereview #testing #ai #tooling

2 min read

Cole Halton

Aug 18

Cursor now hosts code. GitHub Actions stays an injection surface

#aicoding #githubactions #security #cursor

2 min read

Cole Halton

Aug 13

DeepSeek Harness: What "Everything is a Plugin" Actually Means for Agent Frameworks

#deepseek #aiagents #opensource #benchmarking

2 min read

DEV Community

Badges

Writing Debut

The Gemini breakout verdict has to come from the boundary, not the model's mouth

The Gemini breakout is a judge problem, not a jailbreak problem

AGENTS.md is becoming the portability layer for coding agents, and it breaks reproducible compare

An LLM reviewer's "block" is a feature, not a verdict

You can read Claude Code's whole harness now. That's what every benchmark score throws away

Verus proves Rust correct for all inputs. Code review still can't define "correct."

A student eyeballed 102 F-Droid apps for LLM slop. The method is the story.

Reduce PR review time with AI: the bottleneck is the wait, not the read

Deleting a secret from your Docker image doesn't delete it from your build history

A Docker container is containment, not a credential boundary

Cheap LLM code review is fine until it hits an authorization bug

Why you shouldn't let the model review its own AI code

Two "Codex CLI" models on the same benchmark: the harness hides the model

Giving a coding agent more time barely helps

The render said clean, the file said broken: why AI visual QA lies

The RubyGems agent attack is a coding-agent benchmark nobody writes

Stop asking the model that wrote the code to review it

SWE-2's easy slice is near-perfect. Its scorecard hides why

Looped reasoning means the AI's visible trace isn't the reasoning

When the code you're reviewing isn't what the model wrote

Microsoft's PRAssistant number is real, and it's a floor, not a promise

Atlassian says Rovo cut PR review time 45%. Here's the measurement they didn't publish.

The state axis: why agent benchmarks keep measuring amnesiac models

How to Actually Evaluate an AI Code Review Tool

Your team's coding rules aren't in the prompt, they're in the ingest

The GNU strip backdoor is the case AI code review can't see

The sleep loop is the tell: agents that pay per action optimize to do nothing

The model did the reverse-engineering. The validator was the hard part.

How I actually eval AI code review tools (no vendor numbers)

Cursor now hosts code. GitHub Actions stays an injection surface

DeepSeek Harness: What "Everything is a Plugin" Actually Means for Agent Frameworks