Aditya Bhargava argues that the orchestration around a model deserves more attention. He cites a benchmark in which changing harnesses substantially changed scores with the same model, then proposes language-level primitives for tools, interruptions and resumable execution. The benchmark and broader local-model ambition are claims discussed in the talk, not independently established by the demonstrations.
Aditya Bhargava repeatedly uses a faulty median function to show what each addition changes. A model without file tools cannot inspect the problem. Read and write functions add capability, but unrestricted filesystem access is unsafe; interrupt handlers introduce an explicit human decision before sensitive operations.
Aditya Bhargava then fixes tool parameters to a permitted directory using partial function application, reducing repeated approval requests without handing the model arbitrary paths. A reason-act-observe loop runs tests, reads failures, edits code and checks the outcome rather than stopping after suggesting a fix. These mechanisms illustrate scoped behavior in the example, not a blanket guarantee of security.
Aditya Bhargava presents specialized coding and research subagents as callable tools that keep unrelated capabilities out of one context. He also shows an optimizer measuring a baseline and rewriting prompts toward a stated goal. The conclusion is to improve harnesses through observable outcomes, with language support for true pause/resume rather than relying on prompt instructions alone.
Watch on YouTube
