Backtrack to Breakthrough: Why Great AI Agents Revisit
GSM-Agent shows that agent quality depends less on raw search time than on whether the agent knows when to return to promising evidence with a better query.
Cognaptus Insights
Industry thoughts, trends, and commentary.
Lost in the Long Game: What UltraHorizon Reveals About Agent Failure at Scale
UltraHorizon shows why long-horizon AI agents fail less like weak chatbots and more like badly managed investigation teams.
Options = Power: Turning Empowerment into a KPI for AI Agents
EELMA shows how agent optionality can become useful telemetry—but only when we distinguish real control from noisy behavioural diversity.
Paths, Not Parrots: When RL Makes LLMs Plan—and When It Doesn’t
A mechanism-first reading of why reinforcement learning helps LLM planning through exploration, why policy-gradient can collapse into brittle one-path behaviour, and why Q-learning only helps when rewards expose the structure of the task.
Pods over Prompts: Shachi’s Playbook for Serious Agent-Based Simulation
Shachi shows why serious LLM-based agent simulation needs modular cognitive architecture, not just better personas and larger crowds.
Failures, Taxonomized: How Multi‑Level Reflection Turns Agents Into Self‑Learners
How SaMuLe turns failed agent traces into a reusable diagnostic layer—and what that means for enterprise automation.
Paths > Outcomes: Measuring Agent Quality Beyond the Final State
A practical reading of CORE, a path-based evaluation framework showing why tool-using AI agents must be judged by the sequence of actions they take, not only the state they leave behind.
Reason, Reveal, Resist: The Persuasion Duality in Multi‑Agent AI
A mechanism-first look at why reasoning traces can make AI agents both harder to fool and better at fooling each other.
Recon, Then Wreck the Roadblocks: How Recon‑Act Turns Web Stumbles into Tools
Recon-Act shows why browser agents may improve less by clicking harder and more by converting repeated failures into governed, reusable tools.
When Agents Get Bored: Three Baselines Your Autonomy Stack Already Has
A practical reading of what task-free LLM agents do when nobody gives them a job: they build, self-test, or disappear into recursive self-description.