The system
Agents with their own budgets
Each agent gets a budget from the agent that hired it, and spends it hiring models and other agents.
A team of AI agents spends money every time it uses a model or another agent, and a strong cloud model costs more than a cheap one. HASHIRU v2 gives this money to the agents themselves. Each agent gets a budget from the agent that hired it and spends it by hiring models and other agents (over A2A and MCP). HASHIRU's runtime, the hub, keeps this safe. It caps every call at what the agent has left, returns unspent money when a hire is fired, and makes an agent on its last step answer.
The guards
Built-in spending limits
Every call is capped at what its agent has left, unspent money comes back when a hire is fired, and an agent on its last step must answer.
The result
Limits that hold
Across 551 live hires, none spent more than 1.11 times its budget. The first version let one spend 372 times.
01
What the runs show
1.11×
The most any of 551 live hires spent, as a multiple of its budget. The first version let a hire spend 372 times its budget and lost half its runs.
+0.136
Score gained by keeping long documents away from the CEO agent (0.463 to 0.599 of the points, 95% interval +0.067 to +0.216). The CEO's own share of the spending fell from 43% to 32%.
79%vs69%
Right on hard jobs where the market has a specialist (62 of 78), against 54 of 78 for the cheap model alone.
0.737vs0.740
Overall score against the cheap model alone, on projects with enough budget to run the CEO. About the same, at 3.9 times the spend.
What still holds the agents back
The agents pick the expensive model (muse-spark-1.3) for hard jobs, but give it too little money. Its 79 hires took 55% of what the agents spent ($0.99 of $1.81) and got 12 answers right, though it got 87 to 100% of hard jobs right in our test runs. The hub now tells the agent that price when it hires, using numbers it already has. This fix is built and tested, but we have not yet run it live.
02
Splitting the money pays off once the budget has room
We re-scored our recorded results at four budget levels, with no new model calls. The cheap model alone scores 0.740 to 0.749 at every budget, because it has nothing to spend the extra money on. The cascade (try the cheap option first, move up when needed) adapts during the task, as the agents do. It is 0.051 behind the cheap model at the smallest budget and 0.133 ahead at the largest (95% interval +0.078 to +0.192, ahead on 12 of 16 projects).
Show the numbers
| Strategy | 0.25× | 0.5× | 1× | 2× |
|---|---|---|---|---|
| Mean budget a project | $0.036 | $0.072 | $0.144 | $0.288 |
Every job on gpt-6-luna | 0.747 | 0.740 | 0.748 | 0.749 |
Every job on gpt-5-nano | 0.635 | 0.700 | 0.721 | 0.723 |
Every job on muse-spark-1.3 | 0.350 | 0.427 | 0.562 | 0.815 |
| Equal split | 0.748 | 0.761 | 0.791 | 0.808 |
| Cascade (adapts during the task) | 0.696 | 0.712 | 0.783 | 0.882 |
| Knapsack (plans up front) | 0.776 | 0.817 | 0.826 | 0.910 |
| Oracle (knows every answer in advance) | 0.825 | 0.838 | 0.844 | 0.947 |
| HASHIRU v2 agents (live) | 0.737 | |||
Cascade minus gpt-6-luna | −0.051 | −0.028 | +0.035 | +0.133 |
| 95% interval over projects | [−0.108, +0.004] | [−0.088, +0.031] | [−0.036, +0.100] | [+0.078, +0.192] |
03
A small model can learn the split, but not beat the solver
In a third round, we ran open models ourselves on one rented GPU (about 5.5 hours and $24 of credits, with no paid API calls). We trained Qwen3.5 9B to copy the knapsack solver's split of the money from a project's outline, then scored it on all twenty projects of the second round (the chart above uses sixteen), with its own reply paid from the budget.
0.750vs0.788
The trained model against the free knapsack solver, at half of what the best answers cost. It picks the solver's option for 92% of jobs, but its reply costs about 8% of the budget, so it trails the solver at every budget (by 0.023 to 0.052).
+0.17
Gained by training, over the same model untrained (0.578), at that budget. The gain is 0.15 to 0.22 across four budgets, and every 95% interval is above zero. With thinking on, the untrained model never finished, so it answered nothing.
Weaker models did not make the hidden difficulty pay
Before training, we measured Ministral 3 3B, Qwen3.5 9B and Gemma 4 31B. A job's hidden difficulty now moved their success a lot, but knowing it was still worth almost nothing (0.006 of the points), because Gemma 4 31B was both the strongest model and the cheapest per attempt. The trained model is on Hugging Face.
04
Give a team a budget
A team is one YAML file and one command. This file sets up the market from our evaluation, with three cloud models, a paid logic-grid solver reached over A2A, and measured success rates for level-5 arithmetic.
Every request then gets a CEO with a budget of budget_usd, the whole process stops at max_usd, and the log records every payment, call and refund.
# team.yaml ceo_model: gpt-6-luna models: [gpt-5-nano, gpt-6-luna, meta/muse-spark-1.3] budget_usd: 0.10 # the CEO's budget for each request max_usd: 5.00 # real spend, for the whole process request_by_reference: true # long sections stay out of the CEO's context a2a_peers: csp: {url: http://127.0.0.1:9512/} # a priced logic-grid solver success_rates: # kind -> model -> [rate, mean cost in $] arithmetic level 5: {gpt-6-luna: [0.25, 0.0031], gpt-5-nano: [0.53, 0.0073], meta/muse-spark-1.3: [0.92, 0.0961]} $ uv run --env-file .env hashiru ask --config team.yaml "Answer the jobs below ..."
05
When to use HASHIRU v2 today
Five rules of thumb from our runs, each limited to the market we measured.
A hard spending limit
Use HASHIRU when a team must stay within its money. Every hire stayed within 1.11 times its budget, while the same agents without the call cap overspent one by 372 times.
Hard jobs that a specialist can do
Hire through HASHIRU when the market sells exact answers. On hard arithmetic and logic grids, the agents were right 79% of the time, against 69% for the cheap model alone.
A budget too small for the expensive model
Expect the cheap model's score at a higher cost. At the budgets we ran, the agents matched
gpt-6-lunaalone (0.737 against 0.740) at 3.9 times its spend.A very small budget
Below about $0.006 a project, send each job to one cheap model. A CEO that pays for each turn used up the budget before it answered (0.046 against 0.724).
A budget with room
Split the money instead of using one model. In replay, the cheap model alone stayed near 0.75 as the budget grew eightfold while the cascade reached 0.88. We have not yet run the agents at these budgets.
Where it stands
These are small experiments ($33.21 of model calls in total, plus about $24 of GPU time), and our results tables fail three of our seven validity checks, so we cannot yet show what adapting during a task is worth. We have run the agents live only at the smaller budget. The whitepaper gives every interval and the next step for each gap, starting with running the price quote at hire time live and running the agents at larger budgets.
06
Cite
@misc{pai2026agents,
title = {Agents That Hold the Money},
author = {Pai, Kunal and Patel, Harshil},
year = {2026},
howpublished = {{HASHIRU} v2 whitepaper},
url = {https://hashiru-ai.github.io/hashiru-website/}
}