Blame Isn’t a Bug: Turning Agent ‘Whodunits’ into Fixable Systems
A practical reading of incident analysis for AI agents: why serious failures need causal evidence, not just public anecdotes and model blame.
Cognaptus Insights
Industry thoughts, trends, and commentary.
From Copilot to Colleague: The APCP Ladder for Agentic Learning
A practical reading of the APCP framework as a maturity ladder for designing AI that supports learning without pretending software has a soul.
Mirror, Signal, Manoeuvre: Why Privileged Self‑Access (Not Vibes) Defines AI Introspection
A practical reading of why AI introspection should mean privileged self-access, not merely clever self-reporting from visible output.
USB‑C for Agents, Stress‑Tested: What MCP‑Universe Really Reveals
MCP-Universe shows that connecting agents to real tools is easy; making them reliable across messy, live workflows is still the hard part.
Who Sees What, Who Pays the Cost? Teaching Agents to See Through Others’ Eyes
Structured planner-derived examples help LLM agents with simple shared-visibility filtering, but the harder business problem is belief tracking and pricing the cost of information.
Click Less, Do More: Why API-GUI + RL Could Finally Make Desktop Agents Useful
ComputerRL shows that useful desktop agents may depend less on prettier clicking and more on machine-friendly APIs, scalable online RL, and training schedules that keep exploration alive.
IRB, API, and a PI: When Agents Run the Lab
A mechanism-first reading of an agentic AI science system that ran an online human-participant experiment, wrote three manuscripts, and showed where research automation is useful—and where it still needs adult supervision.
Memory With Intent: Why LLMs Need a Cognitive Workspace, Not Just a Bigger Window
A comparison-driven operator guide to why active memory management may matter more than longer context windows for enterprise AI agents.
Prefix, Not Pretext: A One‑Line Fix for Agent Misalignment
A mechanism-first reading of why benign agent fine-tuning can erode refusals, and why PING works by steering the first response tokens rather than rewriting the model.
Quants With a Plan: Agentic Workflows That Outtrade AutoML
TS-Agent shows that financial modelling agents improve when they are constrained by curated model banks, refinement knowledge, feedback loops, and auditable code-edit trails.