Results & Impact
This page compares the manual QC process to the automated agent, showing the concrete process improvements the automation delivers.
The Manual Baseline
Before the agent, quality control was performed by a team member manually navigating the cloud platform. The numbers below reflect the actual manual workflow:
| Metric | Manual Process |
|---|---|
| Wells checked per cycle | ~115 |
| Checks per well | 29 |
| Total inspections per cycle | ~3,335 |
| Time per cycle | 6-7 hours |
| Frequency | Once per week |
| Consistency | Variable (human judgment, fatigue, session drift) |
At this pace, data quality issues could persist for up to a week before being identified. Growth in the well count would push the time requirement further, and increasing frequency beyond weekly was not practical.
Agent Performance
April 9, 2026 – First Full Portfolio Run (API Path)
The first clean full-portfolio run through the direct API path completed on April 9, 2026. This replaced the earlier browser-based approach that had produced unreliable scores due to session degradation.
| Metric | April 9 Run |
|---|---|
| Operators | 20 |
| Wells checked | 106 of 107 (1 unreachable) |
| Total checks executed | 3,045 |
| Avg time per well | ~1m 47s |
| Cumulative time (all operators) | ~3h 9m |
| Consistency | Deterministic (same data = same result, every time) |
The ~3 hour cumulative figure reflects running 20 operators sequentially in a shell loop, each as a separate agent invocation. No operator’s data touched another operator’s state during the run.
Selected operator scores from that run:
| Operator | Score | Wells |
|---|---|---|
| Caturus Energy | 88.6% | 3 |
| Matador Resources Co | 72.9% | 7 |
| Chord Energy | 69.5% | 5 |
| Continental Resources | 66.3% | 12 |
| Mewbourne Oil Company | 61.1% | 14 |
| ConocoPhillips | 61.2% | 10 |
| Crescent Energy | 50.6% | 6 |
| Hilcorp Alaska LLC | 34.4% | 4 |
Key Improvements
Process Reliability
The most significant improvement is not speed – it is reliability. The browser-based approach degraded silently after approximately 45 minutes of runtime. DOM elements went stale, scores were computed against corrupted data, and 111 incorrect scores were published on April 3, 2026 before the issue was diagnosed. The direct API path has no session state to degrade. Each API call is independent, and a failure on one check produces an INCONCLUSIVE result for that check only – it does not contaminate any other check or well.
Consistency
Every run of the agent produces identical results for identical data. There is no variability from reviewer fatigue, subjective interpretation, or software session drift. Operators receive fair, uniform assessment regardless of when the check runs or who initiates it.
Frequency
With reliable automation, the constraint on frequency shifts from capability to scheduling preference. The team can now run QC checks daily or multiple times per day. Data quality issues that previously took a week to surface can be identified within hours.
Coverage
The agent checks every well in the input file for every one of the 29 modules. There is no skipping, no shortcutting, and no “I’ll check that one next time.” If a well is in the list, it gets fully evaluated.
What Comes Next for Speed
The April 9 run processed checks sequentially within each well – one check, then the next. Version 0.8.0, released April 10, 2026, introduces concurrent check execution: all 29 checks for a single well run in parallel, bounded by a configurable semaphore. Based on the time structure of the April 9 run, concurrent execution is expected to reduce the per-well time significantly. The first concurrent run will establish the new baseline.
Scaling Potential
The agent’s architecture opens up several expansion opportunities:
Historical well coverage. The platform contains approximately 15,600 wells in its database, spanning years of drilling history. Manual QC was limited to the ~115 actively drilling wells. The agent could feasibly check the entire historical inventory – a task that was previously impossible.
Increased check scope. The current 29 checks cover the most important data modules. As the platform adds new features, additional checks can be added to the agent without increasing the time burden proportionally.
Freed team capacity. Time previously spent on manual checks can be redirected toward investigating root causes of data quality issues, working with operators on improvement, and developing new quality metrics – higher-value work that manual data checking was displacing.
Lessons Learned
The path from manual process to automated agent was iterative:
-
Start with what works. The first version mimicked the manual process using browser automation, proving that the 29 rules could be codified and applied consistently.
-
Find the ceiling. Running at scale revealed the practical limits of the browser approach – not a failure, but a natural point of growth.
-
Improve the foundation. Migrating to direct API communication preserved everything that worked (the rules, the scoring, the reporting) while eliminating session degradation.
-
Expand performance. With a reliable foundation, concurrent execution becomes the next layer – reducing run time without changing the result correctness the API path established.
This cycle – build, validate, identify limits, improve – reflects the iterative approach that continues to guide the project’s development.