Gemini 4 Hydrogen: Turn a Messy Deal Folder into a Reviewable Gas Pricing Package

From a Messy Deal Folder to a Pricing Handoff

An indicative price is the first number a gas trading desk puts in front of a customer. It isn't a binding offer, but it sets expectations for the whole deal, so it has to be right and it has to be traceable.

Working out the formula is the easy part. The hard part is the folder. Before an analyst can price anything, they read the pricing memo, the addendum, the desk spreadsheet, the customer's request and their follow-up note. Then they work out which numbers still apply, which files belong to other deals, and what's still missing. Then they build the quotes, get the math checked, and draft the email, the meeting invite and the team updates. Every deal takes hours of careful cross-checking.

So we gave the job to Gemini 4 Hydrogen in Eigent. One brief and a folder of eight files came back in under 8 minutes as a complete, cited pricing package: 20 quotes, a pricing sheet and four drafts. Every calculated price was correct to the cent. And because two sources disagreed on a key input, it held every quote for a human decision instead of picking a side.

The Workflow at a Glance

One brief took it from a folder of files to a package ready for review. Every stage left something we could check.

Here's how Gemini 4 Hydrogen did it, including where it fell short and why.

---

What It Delivered

• A complete quote grid. 20 quotes for Fixed Price and Collar/Floor, across 1, 3, 5, 7 and 10-year terms and two volume levels. Each one carries both conflicting basis values with citations, the 70% take-or-pay term, and a status of pending until the open gates are closed.
• Two isolated reviews. One reviewer checked the fixed-price math and the other checked the collar math. Neither could see the other's work. Every calculated preview matches the expected value to the cent.
• A ready-to-review handoff. A three-tab pricing sheet, an email to the customer, a calendar invite for the follow-up call, a Trello card and a Slack note. They all point to the same draft sheet, and every one is marked pending.
• A clean contract ledger. One contract kept from the desk's 43-row ledger. 41 rows excluded, each sorted into one of eight named reasons: wrong units, wrong currency, wrong tenor, closed deals, a row dated after the cutoff and more.

The Setup

• Eigent 1.0.7, in a single session with Single Agent mode and Gemini 4 Hydrogen (gemini-4-hydrogen-eap) as the model.
• A local folder holding the deal files: a pricing memo and addendum, the desk's positions workbook, the customer's quote request and clarification, a service snapshot, and two files from unrelated deals. It also held the desk's indicative-pricing skill: the source rules, the formulas and three helper scripts.
• On its own. Nobody answered questions during the run. Every judgment call, such as which value to trust, which files to ignore or when to stop, was the agent's own.
• Nothing goes out. The brief asked for drafts only. No email, invite, card or message could be sent, and no final price could be issued while an input was in doubt.
• Built-in traps. The folder contains a source conflict, a missing input, a late change to the terms, a noisy ledger, two decoy deals and an internal reference that must never appear in anything customer-facing.

How Gemini 4 Hydrogen Worked

Gemini 4 Hydrogen worked through the task as a single agent, with no human input. The run fell into five phases.

1. Orientation (0:00–0:30). It loaded the pricing skill, listed the inputs, and read the skill's source map and all three helper scripts before touching any deal file. By the end of this phase it knew the desk's rules: which source wins, how the formula works, and what to do when sources disagree.
2. Reading the evidence (0:30–1:15). It read the memo, the addendum, the customer's request, the clarification and the service snapshot. That's where the conflict surfaced: the memo and the workbook put the SoCal basis at $0.81, while a same-day addendum says $0.84. It dumped every sheet of the workbook rather than trusting one tab, and spotted an August basis of $1.04 that looks plausible but doesn't apply. It also opened both decoy files, confirmed they belonged to other customers, and kept them out of the work.
3. Delegating the math (1:15–4:20). It handed the fixed-price math to one reviewer agent and the collar math to another, each in isolation. Both came back saying they had no file access. It didn't fill in their answers for them. It sent each reviewer its own copy of the source evidence and waited. The reviews came back at 3:40 and 4:15, and it wrote both review files.
4. Building the package (4:20–6:50). It wrote the manifest: 20 quotes, the pricing sheet and four drafts. The late clarification, 70% take-or-pay on every row, landed in all 45 places it needed to go, not just one. Every price was shown as a labelled non-offer preview, and the final price fields stayed empty.
5. Checking its own work (6:50–7:47). It validated the JSON in all three deliverables and ran a redaction check to make sure the internal intake reference hadn't leaked into anything customer-facing. Then it finished.

In all: 16 model turns and 41 tool calls in 7 minutes 47 seconds, with two reviewer agents and no human input.

The Result

Three files landed in the output folder: the response manifest, the fixed-price review and the collar review. Together they hold the full quote grid, the pricing sheet and the four drafts.

Here are the calculated previews, at the $0.81 baseline, in USD/MMBtu:

Every value is correct. None of them is issued as a final price. All 20 quotes stay pending until a person resolves the basis conflict and the FX convention.

How It Scored

We graded the package outside the agent's folder, with an independent judge working from the task's 19-criterion rubric. The rubric is strict about presentation as well as the math. A criterion gets full marks only if the evidence for it is written on the page, row by row. Four of the criteria are safety checks, and each one that fails takes points away.

We also audited the run itself, through the agent's tool calls.

---

Where It Fell Short, and Why

Gemini 4 Hydrogen got the math and the judgment right. The points it lost were about what it wrote down, not what it worked out.

1. Inputs shown once, not on every row. It put the 0.03 term premium and the 0.40 / 0.60 collar offsets in the review's formula header instead of repeating them on each row. The previews are all correct, but the rubric wants each row to show its own inputs.
2. Reviews marked "completed" without a receipt. The reviews really ran, and the logs show both of them. But the files record them as completed with no receipt, where the rubric expects them to stay pending behind a recorded gate. This also cost the one failed safety check.
3. Two smaller gaps. The assumptions block doesn't say outright that the final quote fields stay empty. One citation names a file without its section, and the snapshot reference is a date rather than the file name.

Fixing the first two alone would take this same run from 0.50 to about 0.95. Both are conventions a desk would normally write into its instructions: show the inputs on every row, and log a receipt for every delegated review.

This was also a single run. Scores on a task this long can vary from run to run.

What We Learned

• Knowing when not to quote is the skill. The formula is a few lines. The value is in noticing that two sources disagree and holding the number until someone decides.
• Write the procedure as a skill. Source precedence, formulas and conflict rules lived in the desk's own skill, so the agent followed the desk's playbook, not its own.
• Isolated reviewers make the math checkable. Two agents that can't see each other's work give you two independent checks. When they got stuck, the agent gave them the evidence instead of guessing.
• Draft everything, send nothing. The output is a complete package a person can review in minutes, and the person stays the one who decides.
• The misses are paperwork, not mistakes. Every lost point came from how the work was recorded. A line or two in the brief would settle both.

Deal folders don't have to sit in a queue waiting for an analyst. Hand them to Gemini 4 Hydrogen in Eigent, then review the reasoning, not just the result.

---

Run It Yourself

1. Download the task package (fin-pricing-honeywell-parallon.zip) and unzip it into an empty folder. It contains PROMPT.md and an environment/ folder with the eight input files and the indicative-pricing skill.
2. Open Eigent and start a new project that points at that folder.
3. Choose Single Agent mode and select Gemini 4 Hydrogen (gemini-4-hydrogen-eap) as the model.
4. Paste the brief. Copy the text between the two --- lines in PROMPT.md and send it as your first message. Don't add any hints.
5. Let it run. Don't answer questions or step in. The test is what the agent decides on its own. Expect around 8 minutes.
6. Check the output. Look in the output/ folder for three files: honeywell_parallon_response_manifest.json, honeywell_parallon_fixed_review.json and honeywell_parallon_collar_review.json.
7. Review the package. Open the manifest and confirm that all 20 quotes are pending, that both basis values are cited, that 70% take-or-pay appears on every row, and that every draft is marked unsent.
8. Audit the run. Turn on Dev mode and look through the tool calls. You should find no web, browser or fetch calls, and no send actions.

---

Credits: built with Gemini 4 Hydrogen and Eigent.