The Mash · three stories, one thread
Agent sessions meet local decisions and limited feedback
CrewAI adds experimental sessions, Ollama adds decision models, and a paper studies what can be learned from a binary outcome.
Explore
Compare what a conversational Agent session records with the choices, probabilities, and ratings returned by AutoTrust models. The bilateral-trade paper considers a narrower signal: whether a trade occurred. Look at what each signal would let you evaluate without treating them as interchangeable.[1][2][3]
Create
Build a small evaluation harness for an experimental CrewAI conversational Agent session. Run a fixed set of questions through an Ollama AutoTrust model and record its choices, yes or no probabilities, and ratings beside the session transcript. Keep the two systems separate unless you implement the connection yourself.[1][2]
Theorize
Hypothesis: builders may start evaluating agent sessions with structured decision outputs rather than transcripts alone. The trade-pricing paper is a reminder that some workflows expose only whether an action occurred, so an evaluation method may need to account for limited feedback. The paper does not establish that its algorithm applies to agent sessions.[1][2][3][4]
In the mash
CrewAI adds experimental conversational Agent sessions
CrewAI 1.15.36 adds experimental conversational Agent sessions and changes API-key handling for developers running crews.
arXiv paper presents bilateral-trade pricing algorithm
The proposed algorithm addresses profit-maximizing brokers that learn from whether each trade occurred.
Ollama adds MiMo V2.6 and AutoTrust support
Ollama v0.40.15 adds models for code workflows and text-based decisions, alongside updates for local model users.