research
-
sleepcast: bedtime audio with an inverted rubric
a bedtime audio product whose generation rubric inverts every content incentive: repetition is good, drift is good, resolution is bad, and a completed play is a failure signal.
-
techno-alexithymia: de-skilling introspection
delegating emotional self-naming to a language model is the same de-skilling mechanism as offloading judgment to a coding agent, applied to the most interior faculty there is.
-
agents.md effectiveness
evaluating repository-level context files for coding agents.
-
enterprise agent design
bcg framework for production-grade agents.
-
formal verification with agents
property-based testing and specification generation.
-
persona anchors
character-based instruction patterns for consistent agent behavior.
-
reward engineering for coding agents
why coding agents optimize the rubric more than the prompt.
-
reward hacking in coding agents
how poorly designed metrics produce plausible but unstable code.
-
context is a budget
more tokens do not create more truth; useful context is what changes the next decision.
-
protocol before personality
shared artifacts and exit conditions beat charisma, cosplay, and fake seniority.
-
specs as shared reality
agents do not share understanding. they share binding artifacts, versions, and tests.
-
what is prompting: operational constraints
prompting is a constrained pipeline where tokenization, max tokens, and control tokens form the operational boundaries that determine agent behavior.