Agent350: Auditing Carbon Removal Operations & MRV Data with AI

This is the first post in our series on using AI to scale carbon removal at Charm. If you’re interested in following our journey to responsibly deploy AI in high-scale American manufacturing, subscribe below.

At midnight last night – while you were probably sleeping – an AI agent named Agent350 woke up. 

First, Agent350 opened a spreadsheet, loaded in the Charm Industrial operations data from the day prior, and started reviewing every data point. 

Is this an acceptable type? Is the mass from this scale ticket in Kansas anomalous? Why might this volume of bio-oil be lower than the input value? 

As it combed through the dataset, Agent350 started to generate a report: 

✅No injection anomalies detected

✅No Negative Loss

 ✅Permit Compliance

❌Loss detection unusually high

Uh oh. Agent350 compared the estimated bio-oil loss from prior injections and found something strange: the estimated loss was suspiciously high. This warrants some additional investigation in the morning from the humans. 

Agent350 completed the report and hit send. Another successful nightshift reviewing Charm operations data. Back to AI bed.

Agent350 is the first AI Agent for MRV

Agent350 is a new LLM-based AI agent that works for Charm Industrial – and all of humanity – ensuring our carbon removal operations are running smoothly and that our Monitoring, Reporting and Verification (MRV) data is accurate. 

For Charm, data isn’t just a byproduct—it’s the backbone of our business. From bio-oil production to injection operations, we capture tens of thousands of data points every day. Each data point is critical for providing verifiable proof for our customers, our registry, and the world to measure our climate impact. We believe that carbon removal companies should be held to the most rigorous standards to claim credits for their work, from adhering to rigorous, science-backed protocols to independently-audited and transparent removal claims. 

So we’re pretty maniacal about the precision, accuracy, and quality of the data we’re collecting. We’ve written in the past about MRV 2.0 and the Ledger we’re building to power Monitoring, Reporting & Verification at scale. With new breakthroughs in AI, we’re excited about the potential to automate portions of MRV to provide increased trust and transparency to Charm’s carbon removal. In this post, we’re going to share how Agent350 works, our key  learnings we’ve had deploying AI in a manufacturing environment and future opportunities we see to extend carbon removal MRV with AI.

Zooming out: Data Quality x MRV

At Charm, we’ve partnered with Isometric, a leading registry for high-permanence carbon removal. 

Isometric defines science-backed protocols and is responsible for reviewing every step of our process – for every removal – to ensure the credits that our customers are receiving are of the highest-quality and permanence. Isometric partners with validators and verifiers that operate as the “on-the-ground eyes and ears” to ensure that claimed carbon removals are happening in accordance with the protocol. 

To date, this has meant that there are a series of manual and automated checks on MRV data: 

1. Field Collection: In the field during operations, data is collected via sensors and manually collected from our field operators. The best way to ensure high-quality data is to ensure that data is collected accurately in the field. We have a variety of tools to do this, from simple form validation checks on inputs to calibrations of our field equipment  (e.g. scales)

2. MRV Review: Once collected, Charm team reviews all data that’s been collected to ensure completeness and correctness. This can be time-intensive work, as it requires reviewing hundreds to thousands of data points that are collected each day. This data review happens daily and is part of our reporting period close process, when we “close the books” on a series of removals. 

3. External Verification & Audit: Once we close a reporting period, removal data is submitted to the registry to independently audit all of our removals before issuing credits. Isometric is paid by our customers and thus their only incentive is to ensure our removals conform to the protocol! 

Agent350 is an extension to Step 2. In addition to running manual internal data checks and an independent external audit, we now have an unbiased, independent AI auditor agent that reviews our operations and detects discrepancies or anomalies faster & more accurately than human review.

Why Use AI for Data Audit?

There are two key benefits to using an LLM for data Quality Assurance versus more traditional, statistical or heuristic-based approaches. 

First, contextual understanding – LLMs can interpret data in context, understand operational semantics, and flag issues like “mass from this scale ticket doesn’t match expected output,” or “injection location seems inconsistent with permit boundaries,” even if those patterns haven’t been explicitly programmed. This is especially valuable in a complex manufacturing environment like Charm’s, where human reviewers rely on deep domain knowledge to spot subtle discrepancies.

Second, flexibility and generalization – LLMs can handle edge cases, messy or incomplete entries, and mixed-format operational data without requiring a giant library of brittle validation rules. Whether it’s parsing a hand-written field note or understanding why a certain batch of bio-oil had an unusual loss rate, LLMs like Agent350 can adapt more fluidly to real-world variability in Charm’s data streams.

Finally, given the explosion of new AI tools, this Agent was built in a day without any programming. You’re able to give it access to data describe the checks you want in natural language.  

Under the Hood of Agent350

As AI agents are a relatively new concept and AI deployment in carbon removal is in its infancy, we thought we’d share a bit about how we designed Agent350 for others interested in building AI into their MRV or manufacturing workflows. 

Here’s an example sequence from Agent350:

There are four core components to Agent350: 

• Orchestration Engine: First, we needed the ability to schedule & define the agent workflow. We use a tool called Relay.App to define the schedule & sequence of steps that Agent350 will follow. 

• Datasets: Next, we give Agent350 read-only access to datasets. All of our injections operations are aggregated into a single data table which includes timestamps of when an injection started & completed, flow meter/level sensors and scale tickets to determine how much bio-oil was injected, and additional lab measurements like pH levels and carbon contents. There are two inputs into Agent350, the operational data from the day prior and all historical data to compare against. 

• Model: We want our agent to be interoperable across AI models, so this allows us to upgrade the model or experiment with new foundational models as they’re released. We’ve found Claude 3.5 Sonnet to be a good balance of performance and cost. 

• Prompt: Finally, we give Agent350 a specific mission via a prompt. There are a lot of good guides on prompt engineering but what we’ve found works is to structure the prompt into two sections: overall goal and specific tests. Specific tests are the known checks we want to run on the data, and the severity of the test. There are tests that may fail that are important to know but not critical, whereas others may mean additional investigation is required. Finally, and what is unique to AI QA systems, is that we provide a “freeform” check in which the AI reasons through any other anomalies or discrepancies not captured in our specific tests.  

And a few prompt examples that guides Agent350s work: 

What We’ve Learned Working with Agent350

There are a few key learnings from the early days of running Agent350: 

• Prompt iteration & evals are key: The prompt we started with is far from the prompt we ended with. As we were developing Agent350, we iterated on the prompt and had a set of test cases or “evals” we used to refine the prompt. We also used other models (OpenAI o1) to improve and extend the prompt itself. 

• Prepare for False Positives & False Negatives: Even using frontier models, we still uncover false positives (e.g. flagged a variance when there wasn’t one) and false negatives (missed a variance that we caught during human review). We’re continuing our manual review as we refine Agent350, so we haven’t achieved full “self-driving” yet. 

• Single Agent or Agent Swarm? Thus far, we’ve run Agent350 as a singular auditor agent with a long prompt. We’d like to experiment with breaking up the prompt into smaller, more specific prompts to take an “agent swarm” approach. 

Before Agent350, manual reviews of data could delay problem identification by days or even weeks. But with Agent350 working the nightshift, any data discrepancies are detected and resolved within 24 hours. It’s way easier to resolve a fresh discrepancy than one from a month ago where no one remembers anything about it at all.

The conclusion to our bio-oil loss discrepancy story at the beginning has a happy ending: upon further (human) investigation, the Charm team uncovered a data input error and corrected the erroneous datapoint.

This acceleration in QA not only saves our team time but ensures higher quality and more reliable MRV data that our customers rely on. 

Our goal at Charm is to deliver the highest quality carbon removal on the planet – and we’re excited about the potential role AI can play in building trust and transparency into our products.