A.P
Initializing_Core0%
Gathering_Ecosystem_Signals...
ABHINAV PANGARIA
Series
03 parts
25 min total
Since Sep 21, 2026

Your Agent Might Actually Lie to You

Three parts on the hackathon build: a Stop hook that will not let a coding agent say done until a real browser agrees.

I built GuardianKane for the TestMuAI Kane CLI hackathon, where it placed 2nd. This is the story of that first version: why the gate exists, the four paired experiments that tested it, and what I found when I re-read my own logs and had to ask who checks the checker.

Start with part 01
Series cover: Your agent might actually lie to you, with three stacked part tiles: the broken chart, the unasked question, and who checks the checker.
The parts, in reading order
  1. 01
    Part 01 of 03 cover: a line chart drawn twice, slightly offset, under the title My agent shipped a broken chart and told me it was done
    Sep 21, 2026 · 8 min read

    My Agent Shipped a Broken Chart and Told Me It Was Done

    Two builds, one PRD. One rendered a clean performance chart. The other drew the same line twice and reported the task complete. How that gap turned a planning tool into a Claude Code Stop hook that makes the agent prove its work in a real browser, and what the gate caught first.

  2. 02
    Part 02 of 03 cover: a PRD page with a dashed lime box around a missing line, under the title The requirement nobody wrote
    Sep 21, 2026 · 8 min read

    The Requirement Nobody Wrote

    A four-line todo PRD said priority defaults to Medium and nothing else. One build asked how a task becomes High; the other never did. Then I built a trap for the gate and both builds walked straight past it. Four paired experiments, and the pattern they drew: the gap grows with the density of the spec.

  3. 03
    Part 03 of 03 cover: verdict chips reading status failed, confirmed false, and automation_bug, beside a magnifier with a lime question mark, under the title Who verifies the verifier?
    Sep 21, 2026 · 9 min read

    Who Verifies the Verifier?

    On the densest build, eight tasks failed a browser check and one of them was a real bug. The rest were the verifier, and then my own write-up turned them into findings. Notes on checking the thing that checks, and on a three-strike retry cap the agent could quietly reset.