Learn Agentic AI
Articles about Learn Agentic AI
What is a self-scaling benchmark? Coding agents now write the bugs they’re graded on
One agent hides a bug in a real PR, the other must find it blind. Newer models wrote harder tests.
Your agent’s compaction summary keeps losing the details. REMORY adds a second memory made of vectors
Summaries keep the plot and drop the receipts. A learned memory matched full context with 5.2% of the input.
Deleting failed edits is why your agent optimizer stalls
Keeping and refining rejected candidates hit GEPA's AppWorld target with 65.5% fewer rollouts.
Copying your agent’s wins can teach it to loop. Rewriting them first added 17 points
Fine-tuning on raw harness-assisted wins made a 27B agent worse. Rewriting them lifted Terminal-Bench 2 from 57% to 74%.
Self-evolving agents learn to agree on wrong answers. Splitting the data in two fixes it
When the same model writes the answer key and grades the answers, agreement grows faster than accuracy.
What is Belief Trapping? Why your agent needs a state, not a longer memory
Long-horizon agents rarely fail from forgetting. They fail because nothing checks what they believe right now.
More samples won’t fix your agent. A better judge will
A 9B terminal agent jumped from 50% to 68% without retraining. It just started choosing its commands before running them.
In agent distillation, up to 95% of each rollout turn is the student thinking. ActFirst-OPD takes it off the clock
The student acts first and explains itself later. Agent distillation got up to 4.9x faster.
What is Compositional Environment Scaling? How 448 fake apps taught a 35B model to beat frontier models at business workflows
Same data budget, opposite results: single-app fine-tuning made the model worse, composed cross-app tasks tripled it.
The hottest new AI model can’t write a sentence. It just judged 500 agent traces for 18 cents
Jev matched a human reviewer on all 500 agent-trace judgments at $0.00035 a call. Verification just became ambient.
The best thing a world model can predict isn’t the world. It’s your agent’s next mistake
Renmin's AEWM edits an agent's reasoning before it runs. A 9B with the editor beat a 35B without.
Microsoft just measured agent taste. On the hardest decisions, every frontier model scores below random
More reasoning does not fix it. Hindsight distilled from your own agent logs does: 14.6% to 33.7% on SWE-bench Pro.
What is EvoOntology? Why your agent’s data dictionary belongs behind a tool, not in the prompt
The same data dictionary that cut agent accuracy 15 points in the prompt gained 17.8 served as MCP tools
Ant mined a million agent skills from GitHub. No agent ever ran
Code2Skill verifies a million repo-mined skills by rebuilding code from descriptions alone, then beats skills learned from experience