Novoflux published a new stack analysis showing that a model ran the same across three agent harnesses on identical tasks and found 70-fold variation in token use.
Public source
Publisher name
Public post
A New Stack analysis ran the same model across three agent harnesses on identical tasks and found 70-fold variation in token use. Aider used roughly 3,500 tokens per sol…
Company
Novoflux
We Sense so that you can Seize. Reconfiguring capabilities for AI-driven competitive advantage.
- Industry
- IT Services and IT Consulting
- Location
- London, GB
- Company size
- 1 employee
About Novoflux
Novoflux operates at the intersection of AI research and enterprise transformation. SenseLab is our independent research function running frontier experiments on the AI-native transition and publishing findings that corporate technology teams can transfer into their own operations. Hexalink Compass is our structured capability evaluation service using the CAP framework to identify precisely where AI creates value across an entire organisation, and in what order to act. Together, the two services move a company from sensing where AI is going to knowing exactly where to put it.
See moreLatest activity
Latest activity from Novoflux
4 signals
Research & Knowledge
Novoflux published a New Stack analysis that ran the same model through three agent harnesses on identical tasks and documented 70-fold token variation in output.
Research & Knowledge
Novoflux published a New Stack analysis documenting 70-fold token variation across the same model run through three different harnesses on identical tasks, with the difference traced entirely to context overhead, not model behaviour.
Products & Services
Novoflux engineered Compass Agent Build Assist (CABA) engagements with Decryptogen, including compliance gate, independent judge, human in the loop, and full audit trail.
Discover more
Similar signals
Similar public activity from other companies.
Research & Knowledge
xdevops
xdevops improved the routing layer of an agentic system by benchmarking and measuring performance across 1,000 queries.
Research & Knowledge
NewRocket
NewRocket published an article by Melissa Cohoe exploring why agentic AI is forcing the next evolution of cybersecurity and what organizations need to do to maintain visibility, accountability, and control as AI becomes more autonomous.
Research & Knowledge
Appliscale
Appliscale reported that over 27 days of data showed that 3.064B tokens and ~$2,294 estimated API cost were spent on remembering context rather than generating answers.
Research & Knowledge
Infosys
Infosys published a study exploring 26 prompt conditions with AI agents implementing the Zstandard algorithm in Rust.
Research & Knowledge
Workloft.ai