Were you rushing or were you dragging?

18 Mar 2026 · 2 min read

Fletcher: Were you rushing or were you dragging?
Neiman: Rushing
Fletcher: So, you DO know the difference!

To complete one task, an agent does multiple LLM calls. Say each LLM call is compute efficient (adaptive thinking). A not-so-good agent harness with all adaptive-thinking-LLM calls could still yield suboptimal value per token per task. What will elevating this ‘adaptive thinking’ idea to a harness level look like? Some musings:

‘Thinking’ Token Elasticity

Randomly assign tasks to different effort levels. We can then plot value (e.g. accuracy, hours saved, dollars made etc.) against count of thinking tokens. See what’s the optimal effort level where return on thinking token (ROTT, similar to ROAS from paid acquisition world!) is maximized. Run this experiment at some cadence or whenever nature of tasks evolve/drift, new llm models become available etc to determine if your current harness is rushing or dragging? Then auto-evolve the harness based on experiment results.

LLM-Response Apportionment

Analyze end-to-end task traces to see which LLM-Response were the most valuable to achieve the final outcome for that task. In ads, we used a gazillion different apportionment models to attribute the final conversion $$$ to each previous customer touchpoint (email, push notification, search & social ads etc.). This lets your harness know which tools, mcp are providing the maximum value tokens for accomplishing the tasks & invest more in evolving the harness around that boundary.

Context Incrementality

Randomly (& gracefully) turn off tools/skill/mcp/memory etc. to evaluate if that is truly incremental context. If not, remove it.

@article{jain2026were-you-rushing-or-were-you-d,
  title   = {Were you rushing or were you dragging?},
  author  = {Jain, Purak},
  journal = {purakjain.com},
  year    = {2026},
  month   = {Mar},
  url     = {https://purakjain.com/writing/blog/2026/03/18/rushing-or-dragging/}
}