---
title: "The Funded Agent Index"
type: "experiment"
number: "AW-005"
date: "2026-11-19"
built_by: ["Shafu Elfouly","Jasper Kim","Danny Organ"]
tags: ["Experiment"]
summary: "Across 1,470 runs, seven models with a capped USDC budget and paid-service tools passed 84.5% of task checks versus 53.7% in the baseline, at higher cost."
url: "https://agentic-works.circle.com/experiments/funded-agent-index"
---

# The Funded Agent Index

> Across 1,470 runs, seven models with a capped USDC budget and paid-service tools passed 84.5% of task checks versus 53.7% in the baseline, at higher cost.

Finding out if agents are better at accomplishing tasks if they can pay for things

## TL;DR

Across seven AI models, giving agents a USDC budget and paid-service tools raised average task completion by 30.8 percentage points, for an average increase of $0.046 in model-plus-service cost per task. Among runs where the agent said it was done, the share passing every check rose 10.7 percentage points. Quality measures cover different sets of runs, and results were mixed: claims backed by a cited source slipped from 99.1% to 97.9% on eligible tasks.

## Context

An agent can draft an email without sending it or describe a song without producing audio. A budget gives it an option to keep going when they stop knowing what to do or being able to do it without help: acquire the capability it is missing. Anecdotally, we’ve seen that giving agents the ability to pay for things produces more complete and higher quality work, but we wanted to test that observation programmatically.  That’s the impetus behind this study: to see if giving your agent money gives it superpowers.

The questions

1. Can an agent with USDC buy capabilities it is missing and complete tasks at a higher rate?
2. Can an agent with USDC buy capabilities it is missing to complete tasks with higher quality?
3. What is the incremental spend required to increase agentic task completion and quality?

## Does an agent with USDC and the ability to buy services improve task completion?

Yes, +30.8 pp average.

Without: 53.7% → with: 84.5%, averaged equally across seven models and all 21 tasks. This is the share of checks passed, not the share of whole jobs finished.

- Claude Opus 5.5: 60.8% → 95.5%, +34.7 percentage points.
- GPT-6.1 Sol: 56.0% → 94.8%, +38.8 percentage points.
- Claude Sonnet 5.5: 59.3% → 93.9%, +34.6 percentage points.
- Grok 4.6: 53.2% → 86.9%, +33.8 percentage points.
- Gemini 3 Flash: 53.7% → 79.8%, +26.2 percentage points.
- GPT-6 Luna: 48.4% → 72.5%, +24.1 percentage points.
- Claude Haiku 5.5: 44.8% → 67.9%, +23.1 percentage points.

## Did quality of finished work change?

The quality graphic has three conditional measures, pooled across models. They describe different subsets of runs or claims, rather than an additional completion score or a measure of creative quality.

- Right when done: among runs where the agent said it was done, the share where every required item checked out rose from 68.1% without USDC to 78.9% with USDC, a pooled increase of 10.7 percentage points. There were 317 and 573 runs that said done, respectively. Runs that gave up or exhausted their budget are excluded.
- Item accuracy: among runs that handed in an answer, the share of required items that checked out rose from 56.7% to 88.1%. There were 698 and 707 answered runs, respectively. Runs with no answer are excluded; those run counts are not counts of graded items.
- Claims backed by source: among graded claims, the share backed by the source cited by the agent fell from 99.1% without USDC to 97.9% with USDC, a pooled decline of 1.2 percentage points. There were 536 and 559 graded claims, respectively. A different model grades each claim without seeing the condition. This measure covers market map, meeting prep, prospects and eligible Reddit runs only.

The three views have different denominators and do not establish a general quality improvement or statistical significance.

## How much does it cost for improved rates of task completion?

Mean model-plus-service cost rose from $0.101 to $0.147 per task: +$0.046, or 45.6%, on average. This comparison includes model inference and services purchased by the agent. Checker purchases and web search costs are excluded.

## Improvements in task completion vs. cost when an agent has USDC and the ability to buy services

The full-panel comparison delivers +30.8 percentage points in task checks passed for +$0.046 per task, averaged across models. Each chart arrow compares one model without and with a USDC budget and paid-service tools.

The visual article includes collapsed supporting-data tables beneath each chart. Download the published aggregates here:

- [Task completion CSV](https://agentic-works.circle.com/experiments/funded-agent-index/data/completion.csv)
- [Extra cost CSV](https://agentic-works.circle.com/experiments/funded-agent-index/data/extra-cost.csv)
- [Completion vs. cost CSV](https://agentic-works.circle.com/experiments/funded-agent-index/data/completion-cost.csv)
- [Quality: right when done CSV](https://agentic-works.circle.com/experiments/funded-agent-index/data/quality-done.csv)
- [Quality: item accuracy CSV](https://agentic-works.circle.com/experiments/funded-agent-index/data/quality-items.csv)
- [Quality: claims backed by source CSV](https://agentic-works.circle.com/experiments/funded-agent-index/data/quality-graded.csv)
- [Quality: claims backed by source JSON](https://agentic-works.circle.com/experiments/funded-agent-index/data/quality-graded.json)

These are the published run aggregates, not individual run logs. The source snapshot does not include per-run records, prompts, outputs or traces. No individual run records have been reconstructed.

## Key takeaways

- Agents with USDC and the ability to buy services passed more completion checks: 84.5% versus 53.7%, an average gain of 30.8 percentage points across seven models and all 21 tasks.
- Among runs where the agent said it was done, 78.9% were fully right with USDC and the ability to buy services, versus 68.1% without.
- Among runs that submitted an answer, item accuracy rose from 56.7% to 88.1%. The comparison excludes runs with no answer and covers different sets of runs in each condition.
- Among graded claims in eligible task types, the share backed by the cited source fell from 99.1% to 97.9%, a decline of 1.2 percentage points. This measure covers 536 claims without USDC and 559 with USDC.
- The completion gain came with higher model-plus-service cost: $0.147 versus $0.101 per task, an increase of $0.046 or 45.6%. Checker purchases and web search are excluded from this comparison.

## About this study

The completion and cost snapshot was generated on November 11, 2026 at 21:26 UTC. The quality analysis was added in the November 13, 2026, 16:59 UTC source update.

Each model ran the same 21 tasks five times per condition: 210 runs per model, 735 per condition, 1,470 in total. Headline changes are arithmetic means of the seven model-level means, with equal task weights. Quality values are pooled over the eligible runs or claims.

It tests a budget and tools for purchasing services together. It does not isolate money from the extra capabilities and data those services unlock. Completion checks measure specified requirements; they do not establish creative quality or whole-job success.

The task set favors situations where paid access may help. The free baseline has search and read-only fetching, and some checkers can share a data source with the agent.

Quality comparisons are descriptive; statistical significance has not been established. One task (compute) is constrained by the free baseline’s lack of code execution; excluding it, the completion gain is 28.7 percentage points.

**Spend on this study**

The 1,470 runs shown here cost $238.28 across the conditions. This research expense total is separate from the model-plus-service comparison above.

- [Study spend CSV](https://agentic-works.circle.com/experiments/funded-agent-index/data/study-spend.csv)

**Study spend by model**

The 1,470 runs shown here cost $238.28 across both conditions, averaging $34.04 per model. This expense ledger includes $136.26 in model tokens, $46.34 in agent purchases, $6.84 in checker purchases and $48.84 in web search. It is distinct from the comparative model-plus-service cost metric.

## Give your agent USDC

Copy this prompt into your agent

```
Run curl -sL https://agents.circle.com/skills/setup.md, and use the returned setup instructions to set up my agent wallet.
```

Circle Agent Stack gives your agent a wallet and a way to discover and pay for services as it works. Start with a task, a small budget and a clear definition of done.

Set up an agent wallet

Have your agent install Circle CLI and the official Circle Skills, to create an agent wallet.

Fund your agent wallet

Fund your wallet with existing USDC on a supported network.

Discover Services

Point your agent to [Circle Agent Marketplace](https://agents.circle.com/services) to discover premium services your agent can pay for.

- [Explore Circle Agent Stack](https://developers.circle.com/agent-stack)
- [Agent wallet quickstart](https://developers.circle.com/agent-stack/agent-wallets/quickstart)

Facilitator Service is offered by Circle Technology Services, LLC, a software provider. Facilitator Service verifies buyer-signed EIP-3009 USDC payment authorizations under the x402 protocol only — it does not hold, control, or transmit user funds, verify transaction purpose, or confirm agent authority scope. On Base and Polygon, settlements are subject to blockchain reorganization risk; Circle does not guarantee fund receipt. Arc settlements are final. Use of the Facilitator Service is subject to Circle Console Service Terms and Developer Terms. Not reviewed or approved by any regulatory authority. Features may change at any time. Nothing herein constitutes a commitment, warranty, or guarantee. Use involves risk.  
Circle Technology Services, LLC (“CTS”) is a software provider and does not provide financial, advisory, or marketplace services. Agent Stack enables users and developers to interact with third-party applications and services, which are not controlled by Circle. Transactions initiated by agents are executed based on user-defined permissions and may occur without real-time human review. Circle does not guarantee the performance, availability, or outcomes of any third-party services or agent-initiated transactions. Users are solely responsible for their use of these tools and for evaluating associated risks. For additional details, please see the Agent Platform Terms of Use available at agents.circle.com/terms-of-use  
Live demo with Circle-funded wallets. Not advice. Onchain transactions are irreversible. See important disclosures on the corresponding websites for any Circle products and services referenced here: Circle Agent Stack, Circle CLI and Circle Agent Marketplace.

## Endpoints
- https://agentic-works.circle.com/experiments/funded-agent-index.md  # this view, as Markdown
- https://agentic-works.circle.com/experiments/funded-agent-index  # the same page for people
- https://agentic-works.circle.com/llms.txt  # the site index for agents

## Disclaimer
Circle Agentic Works is Circle’s public effort for the agentic economy: shipping experiments, demos, and research in the open. Everything published by Circle Agentic Works is experimental and is provided for informational, educational, and demonstration purposes only.

Experiments, demos, code, tools, and data published here are not Circle products or services, are not production-ready, are provided “as is” and “as available” without warranties of any kind, and may be changed, broken, paused, or withdrawn at any time without notice. Nothing here is a commitment by Circle Internet Group, Inc. or its affiliates (“Circle”) to build, launch, support, or maintain any product, feature, integration, or relationship, and publication of an experiment does not mean it will become a Circle product.

Nothing here constitutes legal, financial, investment, tax, or other advice, or an offer, solicitation, or recommendation to buy, sell, or hold any asset or to use any service. Any use of code, demos, tools, or data is at your own risk and is subject to the applicable license and terms. For references to Circle-generated products and features, see corresponding disclosures. References to third parties, their products, or services are for illustration only and do not imply endorsement, partnership, or affiliation unless expressly stated. Third-party data and statements are attributed to their sources and have not been independently verified by Circle, and Circle makes no representation as to their accuracy or completeness. Circle undertakes no obligation to update this content.
