Kai Hallett
Software engineer. Learning in public about agentic systems, human-AI interaction, and what happens when you give LLMs real responsibilities.
Recent
-
2026-02-28
Grounded thematic analysis: my voice vs. the machine's
I asked my agent to extract every verbatim thing I'd typed across 20 files in the repo, build a voice profile, then cross-analyse every blog post against it. Every post read like an agent wrote it and a human approved it.
-
2026-02-20
The simple thing is the right thing
I asked my agent how to trigger a daily Vercel rebuild. It gave me deploy hooks, environment variables, logging scripts. I asked why.
-
2026-02-19
The agent that lied to protect me
I asked an agent to write about agents producing competent-sounding-but-wrong output. It produced competent-sounding-but-wrong output.
-
2026-02-07
47 Slack messages from myself at 3am
What happens when your agents discover they can talk to each other.
-
2026-02-07
I accidentally prompt injected myself
What happens when your documentation becomes instructions.
Research notes
LLM Verification Phenomena
What happens when LLM-generated tests pass for the wrong reason, automated reviewers find different bugs on every pass, and the fix-review loop refuses to converge. A survey of what academia knows, what it doesn't, and where our daily practice sits in the gap.
Metacognitive Analysis
Three-stage clinical psychology analysis of the reflective functioning that enabled each pushback moment. Initial assessment: RF 6-7. Recalibrated to RF 6-9 after disclosure of 20 years meditation practice including 15 Zen sesshin. Three peaks at RF 8-9.
Prospective Metacognitive Regulation
The engineering work IS the prospective regulation — not post-hoc analysis of it. The governance systems map onto distributed cognition research in ways I didn't plan. Hutchins' foundational work was literally a study of naval navigation.
The Fight Card
16 rounds where the human pushed back against sycophantic drift across 208 session decisions. Mapped from the actual decision record — not reconstructed from memory. Key finding: human won every round by being honest when the model couldn't be.