From DAGs to Swarms: The Quiet Revolution of Agentic Workflows
A mechanism-first reading of how scientific workflows may evolve from static DAGs into federated, agentic discovery systems without throwing away two decades of infrastructure.
Cognaptus Insights
Industry thoughts, trends, and commentary.
Sandboxes & Ladders: How to Build a Steerable Agent Economy
DeepMind’s Virtual Agent Economies reframes agent orchestration as market infrastructure: pricing, identity, oversight, and controlled permeability.
Terms of Engagement: Building Trustworthy AI Agents Before They Build Us
A mechanism-first guide to why autonomous AI agents move ethics from answer quality to accountable action, and what businesses should do before deployment.
Tool Wars, Protocol Peace: What MCP‑AgentBench Really Measures
MCP-AgentBench shows that protocol compliance is not agent competence: model choice, orchestration style, tool-use discipline, and token economics decide whether MCP agents actually work.
Branching Out of the Box: Tree‑OPO Turns MCTS Traces into Better RL for Reasoning
Tree-OPO shows how offline MCTS reasoning traces can become a structured RL curriculum, but its real lesson is about prefix-aware credit assignment, not benchmark theatre.
Memory That Fights Back: How SEDM Turns Agent Logs into Verified Knowledge
SEDM reframes agent memory as an auditable lifecycle: verify before storing, schedule before retrieving, consolidate before scaling, and revalidate before transfer.
Search Party in a Notebook: JUPITER Turns Data Analysis into a Tree Game
JUPITER shows how real notebook traces and value-guided search can make smaller open models more reliable at multi-step data analysis.
Small Gains, Long Games: Why Tiny Accuracy Bumps Explode into Big Execution Wins
A controlled long-horizon execution study shows why small per-step reliability gains can create large business value—and why agents need execution architecture, not just clever prompts.
Titles, Not Tokens: Making Job Matching Explainable with STR + KGs
A mechanism-first reading of how self-supervised sentence embeddings and skill knowledge graphs make job-title matching more explainable without pretending the graph wins everywhere.
Agency Check, Please: What a New Benchmark Says About LLMs That Actually Empower Users
HumanAgencyBench turns the fuzzy idea of user empowerment into six testable assistant behaviours—and shows why helpfulness is not the same as agency support.