AI Explained
Plain explanations of trending AI concepts, with live visualizations.
Scaling Discovery Requires Forgetting — Periodic restart from the best candidate — What does it mean?
Discovery agents plateau as history piles up; wiping it and restarting from the best candidate lets the same budget go further.
Cohere coding-agent teams — Planning-only delegation — What does it mean?
Delegating only planning to a coding-agent team matched full delegation's reward at half the cost.
Knowledge Graphs as Files (DKGI) — agent-written graph tools — What does it mean?
Agents that write Python over a graph's raw files scored 75.2 Hit@1 against 72.8 for fixed graph tools in a controlled test.
Regularizing agents across renamed tool interfaces — CIPR consistency KL — What does it mean?
A consistency KL makes a tool-calling agent pick the same tool whatever the tool is named or where it sits.
BOTTLED — Amortizing an agent into a reusable artifact — What does it mean?
Agents must turn a big model's skill into a cheap reusable artifact; zero-shot scores barely predict who can.
Tev-Gym — Progress-potential reward shaping vs terminal-only reward — What does it mean?
A progress score lets failed agent rollouts be ranked, so GRPO still learns when nobody passes.
RATED reranks agent tools by live reliability — Reliability-aware tool reranking — What does it mean?
RATED shortlists tools by meaning, then reranks them with a fast-forgetting bandit that learns which tools actually work right now.
BLADE audits multi-agent handoffs to localize failure — Handoff auditing — What does it mean?
To find which agent broke a failed team run, BLADE audits how each agent used its input, not just who fed whom.
POrchestra lets subagents ask for harness changes mid-run — Mid-run subagent harness requests vs fixed briefs — What does it mean?
A subagent can pause, ask the orchestrator for a different goal, tools, context or budget, and resume with its work intact.
Microsoft releases Decision-1 for single-pass agent control — Calibrated option probabilities — What does it mean?
A 0.9 should be right 9 times in 10. Decision-1 scores fixed options in one pass so agents can gate on calibrated confidence.
PASAC: lessons from three 2026 agent-eval breaches — Continuous boundary assurance — What does it mean?
An agent's sandbox boundary is a claim to measure and enforce while it runs, not a setting to trust.
X-AgentRed — Tool-dependency graphs for harmful paths — What does it mean?
X-AgentRed maps which tool's output unlocks which tool's input, then walks that graph backward from a harm to find risky agent paths.