A coding agent’s harness can triple its cost without making it more accurate, study finds
Researchers held five models fixed across three production harnesses. A vendor’s own harness bought no edge, and the only effect that beat the noise was a harness losing tasks.
Ben Dickson
Researchers made AI agents better by turning academic papers into working code — no retraining required
ScholarEvolve finds new ways to improve AI agents by turning academic research into code, boosting performance without retraining the underlying model.
Ben Dickson
A new framework teaches AI models to audit and repair each other’s reasoning, improving accuracy
Standard multi-agent pipelines break down when complex problems cannot be divided into subtasks in advance. Self-Organizing Agent Teams (SAT) solve this by learning reusable collaboration strategies from past transcripts.
Ben Dickson
H-JEPA teaches world models to plan at multiple levels of abstraction
Researchers at Yann LeCun’s Advanced Machine Intelligence (AMI) Labs and academic institutions have developed H-JEPA, a world-model architecture that learns to plan at different timescales using separate representations of the environment.
Ben Dickson
Why giving AI agents more compute isn't enough—and what Meta proposes instead
AI agents often struggle to assess their own progress and decide what to do next. They can waste compute resources following failed approaches, or they might overlook promising results or stop before completing a task. These problems become more costly as agents take on longer and more complex workflows.
Ben Dickson
New MIT and Sakana AI framework uses an LLM judge to cut evaluation costs for self-improving coding agents
MIT's SIFT framework uses a language model to cut coding agent evaluation costs, achieving 35.1% accuracy on Polyglot with reduced compute resources.
Ben Dickson
Google’s WikiSkill gives AI agents a memory of what went wrong — without putting it in the prompt
In tests, WikiSkill helped a 9B Qwen outscore a 27B model by turning past runs into reusable skills.
Ben Dickson
Stanford and Nvidia's open CLM-8B caches reusable agent actions and runs up to 9x faster than Jev in tests
Instead of writing out an answer, the open 8B model scores a fixed list of options. It ran faster than TypeSafe's Jev in the team's tests but gave up a few points of accuracy on tool calling.
Ben Dickson
Text handoffs slow AI models down. C2C lets them communicate through KV caches instead
C2C cuts out text handoffs between models — but only for teams that control their own inference stack.
Ben Dickson
Google’s open source EnvHarness lets AI agents train against environments that evolve with them
The approach presents an alternative to continuously building new simulators and training tasks from scratch: start with a trusted environment and dynamically reshape it around the agent’s current weaknesses.
Ben Dickson
Salesforce researchers took an AI agent from finishing 43.5% of browser tasks to 93% without touching the model
Fixing one agent failure tends to break another. Salesforce's answer is to evolve rival versions of the harness and promote only the ones that don't regress.
Ben Dickson
Frontier models can recover up to 65% of facts they can't directly recall — just by thinking longer
Google Research shows LLMs encode 95-98% of facts, with recall as the main bottleneck, aiding reliable app development without larger models.
Ben Dickson