A customer request can touch several departments and applications before it is resolved. Your people keep it moving: finding the right record, checking what another team has done, and making the next update. Giving an agent access to one part of that process leaves much of the coordination with the employee.
We’re building Monarch to take on that work across the oldest, hardest systems in an enterprise. It connects the actions available in those systems and gives agents the application knowledge and checks they need to carry a process through.
Our latest benchmark tests the execution part of that promise across 150 simulated business workflows from AutomationBench. Monarch’s strongest configuration completed 86% of them. Every Monarch configuration tested completed more workflows than every alternative in the evaluation.
More work completed from beginning to end
We counted a workflow as complete only if the agent finished every requested task without making any unrequested changes.
When an agent misses an update, a customer request can sit unresolved until someone notices. If it changes the wrong record, another team may act on incorrect information. In either case, employees have to work out what happened and finish the job.
We evaluated workflows across finance, HR, marketing, operations, sales, and support. The chart below shows the full comparison, including the model and reasoning setting for each system.
Every Monarch configuration led the alternatives tested
Full workflow completion · All 150 tasks per configuration
That advantage appeared in every department in the same-model comparison. The breakdown below shows how the overall result was distributed across the six functions.
Each department contains 25 workflows, so the breakdown is a guide to where teams should investigate further, rather than a prediction for every process in that function.
Monarch’s lead across each department
GPT-6.1 Sol · High reasoning · 25 workflows per department
What helps Monarch finish the work
Finishing a workflow requires more than choosing the right tool. The agent needs to know how the application behaves and whether its actions produced the intended changes. That is where we have focused the engineering around the model.
Monarch’s Product Graph maps the actions available across connected applications, including legacy systems without clean public APIs. It retains how those applications behave: how records are matched, which fields are stored, and whether repeating an action creates another record.
Monarch uses that knowledge to check for existing records before creating duplicates and to verify changes where the application provides a verification route. It reads back stored fields against the requested values. When a write’s outcome is uncertain, it calls for checking the state before another change is attempted.
During a long workflow, Monarch keeps a record of what the agent has read, which changes are planned, and which requested outcomes have evidence of completion.
Checking before continuing helps avoid duplicate records and catches updates that did not land. Teams can then direct human review to the decisions and exceptions that need it.
Our broader approach combines AI reasoning with deterministic code where the steps can be specified and tested. A repeatable operation should not need to be reasoned through from scratch on every run. Tested code lets us repeat those steps without spending model reasoning on them each time. Translation: more bang for your buck. 🔥
What it costs to run that work repeatedly
Once a workflow can finish correctly, the next question is what it takes to run it repeatedly. More reasoning can help with difficult work, but it also adds time and model cost. Teams need to see whether the additional completions justify that spend.
Within Monarch, reducing the reasoning level shortened the median run and lowered model cost per completion, while leaving three additional workflows unfinished. Teams can use that comparison to decide which setting to test on their own work.
Shorter runs and lower model cost, with three fewer completions
Two reasoning settings within Monarch · GPT-6.1 Sol · 150 workflows each
Workflows completed
3 fewer completions at Medium
Median time per workflow
37.5% shorter median run at Medium
Model cost per completion
About 19% lower at Medium
Monarch completed 126 workflows to Codex’s 76, at roughly the same model cost per completion: $0.29. Both used GPT-6.1 Sol, with Medium reasoning for Monarch and High for Codex. These costs include failed attempts.
Cost per completed workflow
Recorded model usage, including unsuccessful attempts · Lower is better
Put Monarch to work on a process your business depends on
The benchmark shows that more work can finish with the models available today. The next test is a process your business depends on, in the systems your people actually use.
Choose one where employees regularly have to coordinate across applications or correct what an agent has done. Define what a completed job requires, which changes are permitted, and which decisions need human review. Then measure completion, time, and cost, including the work still left for your team.
Our strongest benchmark configuration still left 21 workflows unfinished. In a deployment, those failures and exceptions need to be understood before the business relies on the process. That is the work of turning a promising result into something a team can use repeatedly.
Your business gets value when the work is finished. With the same model at the same reasoning level, Monarch completed 53 more workflows than Codex in this benchmark. Bring that standard to your own operations: more work completed, a cost you can justify, and less left for your team to finish.
Methodology
All nine configurations ran 150 simulated workflows from AutomationBench, split equally across six departments and covering easy, medium, and hard tasks. AutomationBench supplied the task grades. The baseline systems used AutomationBench’s standard tools; Monarch used its own execution system, including the Product Graph. Every task remained in the denominator. Full completion requires every requested outcome to pass with no unrequested changes.
The headline compares GPT-6.1 Sol at High in Monarch and Codex: 129 versus 76 completions. Relative improvement is (129 − 76) ÷ 76 = 69.7%. The completion-rate difference is 35.3 percentage points.
The benchmark measures workflow execution in a simulated setting. It does not separately measure safety improvements, validate every legacy-system capability, or attribute the results to deterministic execution. The broader product approach described in this article is distinct from those measured results.
These are complete-system comparisons, not feature ablations. Some configurations use different models or reasoning settings. The results do not establish statistical significance or guarantee production performance.
Cost per completed workflow is total recorded model cost divided by completed workflows. Monarch Opus High usage remains partly unreconciled. Platform, implementation, and other operating costs are excluded.
This is a different evaluation from the first published benchmark. Its 86% result should not be read as a direct before-and-after comparison with the earlier score.
Bring us a workflow your current agents can’t finish
Choose a process that crosses your systems and leaves your team checking, correcting, or completing the work. We’ll work with you to define what success requires and evaluate Monarch against it.
See Monarch on your workflows