# FinOps for AI Agents: Who Spent All the Tokens? — Tisha Chawla & Susheem Koul, Microsoft

Source: [AI Engineer](https://www.youtube.com/watch?v=GJX19pNhmSw)  
Feed7 permalink: https://feed7.dev/p/finops-for-ai-agents-who-spent-all-the-tokens-tisha-chawla-susheem-koul-0artw9n  
Published: 2026-08-22T14:30:07.000Z  
Trust: Source Linked (source_linked)

## Why Included

Run-level policies cut benchmarked agent spend while completing more runs, suggesting builders should steer loops, context, and tool output before enforcing hard caps.

## Source Summary

Microsoft’s TokenOps prototype governs whole agent runs rather than isolated model requests. Across Browser Use and MetaGPT benchmarks, its full policy suite cut **average spend by almost 78%** and raised completion from **67% to roughly 96%** versus simple throttling.

## Practical Implication

Instrument boundaries between agents and tools, attribute costs per run, and enforce cumulative budgets in the call path. Prefer steering actions such as context compaction and tool-output reduction before halting a run.

## Agent-Ready Context

Microsoft’s TokenOps prototype governs whole agent runs rather than isolated model requests. Across Browser Use and MetaGPT benchmarks, its full policy suite cut **average spend by almost 78%** and raised completion from **67% to roughly 96%** versus simple throttling.

Instrument boundaries between agents and tools, attribute costs per run, and enforce cumulative budgets in the call path. Prefer steering actions such as context compaction and tool-output reduction before halting a run.

The figures come from two open-source repositories and unspecified mixtures of stress, simple, and hard scenarios. The proposed self-learning policy module is an intended future state, not a demonstrated capability.

## Connected Context

Feed7 judgment across 545 accumulated Signals:

This moves observability and control from individual calls to the economics of an entire agent run, adding cumulative cost as a reliability constraint. The reported benchmark suggests that steering context and tool output can improve both spend and completion relative to simple throttling, but the limited scenario disclosure keeps the result provisional and does not establish a self-learning governor.

- [How Forward Deployed Engineering is done at Factory — Eno Reyes](https://feed7.dev/p/how-forward-deployed-engineering-is-done-at-factory-eno-reyes-0zgscmd) — Factory ties autonomy to instrumented workflows and measurable outcomes; TokenOps adds per-run cost attribution and cumulative budgets to that operating model.
- [Agent Frameworks Considered Harmful — Rémi Louf, .txt](https://feed7.dev/p/agent-frameworks-considered-harmful-remi-louf-txt-1ku9s0h) — Durable causal traces and exact request identity provide the runtime evidence needed to attribute spending across agents, tools, and attempts.
- [Evolution of agentic surfaces — Gagan Bhat & Isabella Kai He, Anthropic](https://feed7.dev/p/evolution-of-agentic-surfaces-gagan-bhat-isabella-kai-he-anthropic-0mtshqh) — TokenOps uses compaction as a steering action, while Anthropic’s evidence warns that fixed context interventions can change value or become harmful across model upgrades.

## Context Map

- Layer: agent
- Domains: coding
- Topics: harness-engineering, observability, agent-reliability

## Uncertainty

- The figures come from two open-source repositories and unspecified mixtures of stress, simple, and hard scenarios. The proposed self-learning policy module is an intended future state, not a demonstrated capability.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.