30 Dollars a Week: How a Self-Improving Agent Raised Enterprise Back-Office Accuracy from 41% to 91%

Title

30 Dollars a Week: How a Self-Improving Agent Raised Enterprise Back-Office Accuracy from 41% to 91%

Meta Description

Brainbase Labs deployed a self-improving browser agent for Domino's, replacing visual reasoning with hardcoded coordinates and swapping flagship models for Sonnet 5—raising task accuracy from 41% to 91% and cutting operating costs to $30 per week.

Article Content

Enterprise AI engineering has long operated on a default assumption: reliability comes at a high cost. Improving accuracy typically means larger models, endless retry loops, expensive vision processing, and human fallback teams. But Milind Shah, engineering lead at Brainbase Labs, used the case of Domino's Pizza to prove that this cost curve can be completely inverted.

A Counterintuitive Case: Doubling Accuracy, Slashing Costs to a Fraction

Domino's Pizza had previously relied on an external consulting team to manually track competitor pricing and promotional discounts at local fast-food chains. This approach was not only costly but difficult to scale. Brainbase deployed autonomous browser agents to take over the process.

After nine iterations, the system's task accuracy jumped from 41% to 91%, while operating costs plummeted from thousands of dollars per month in consulting fees to just $20–30 per week—with inference cost per data source as low as $1.16.

Behind this 50-percentage-point accuracy leap lies a methodology of pragmatic AI engineering, not an arms race of model scale.

Phase 1: Define "Correct" Before Writing Code

Before touching any prompts or model parameters, the Brainbase team solved a foundational problem: what does success actually look like?

Domino's had previously relied on human analysts to manually check competitor websites, collect local coupon codes and promotional prices, and build structured spreadsheets for executive decision-making. To replicate this process in software, Shah sat down directly with Domino's non-technical business team and went through the reference spreadsheets they used, item by item.

"They already knew what a good spreadsheet looked like," Shah explained. "We used those to create an internal benchmark dataset with questions and expected outputs. Without an internal benchmark, you can neither prove accuracy numbers nor tie improvements to cost reductions."

This benchmark became the scoring basis for every subsequent prompt adjustment, model swap, and agent instruction change.

Phase 2: Hardcoded Coordinates and Model Optimization

Once the benchmark was established, Brainbase began systematically attacking the agent's execution bottlenecks. Unexpectedly, the biggest accuracy gains didn't come from a larger LLM—they came from eliminating visual reasoning.

In early iterations, the browser agent relied on visual processing to locate buttons and navigation links on competitor websites. This was slow, expensive, and prone to misclicks. Shah realized that fast-food ordering interfaces have static UI layouts. Rather than having the model parse image tokens on every page load to find the "View Deals" button, Brainbase simply hardcoded fixed X/Y screen coordinates for recurring UI elements.

"Instead of having the agent read images to figure out where the button is, we just told it the exact X/Y coordinates," Shah said. "This eliminated expensive image-reading inference, and that single change delivered the biggest accuracy improvement."

That said, Shah also experimented with model selection. Rather than defaulting to Anthropic's flagship model, the team found that highly structured, tightly specified prompts paired with Sonnet 5 delivered better accuracy—at a fraction of the token cost of the flagship model.

The cost trap of visual reasoning: In multimodal agents, visual reasoning refers to the mechanism where a model converts page screenshots into image tokens and parses them frame by frame to determine the positions of clickable elements. Its cost structure differs significantly from text reasoning—a single high-resolution screenshot often consumes hundreds to over a thousand image tokens, and every page state change requires re-inference. More critically, visual localization is stochastic: even logically identical buttons can be misidentified when pixels shift slightly or rendering differs. For websites with highly stable UI layouts and enumerable interaction paths (such as chain fast-food ordering, government forms, or enterprise intranet systems), this "re-read the image every time" approach is a textbook case of repetitive waste. Hardcoding X/Y coordinates is equivalent to downgrading "localization" from a reasoning task to a lookup operation—saving not just token costs but also latency and failure retries.

Phase 3: The Frontier Self-Improvement Hasn't Conquered

Brainbase's platform currently powers over 10,000 enterprise internal agents, covering back-office workflows from invoice anomaly cleanup to inventory reallocation. While Brainbase's automated "Hillclimbing" engine can automatically detect and fix failure modes such as incorrect tool calls or hallucinations, Shah remains candid about the current boundaries of automation.

Complex browser navigation trajectories still require human engineer intervention. "Automated improvement works well for tool calls and hallucinations," Shah noted. "But browser navigation is still too complex for self-improving agents. Engineers still need to manually review agent trajectories to find where browser operations break down."

For teams with tight evaluation budgets, Shah offered straightforward advice: don't overcomplicate your initial evaluation. "Deterministic evaluation criteria go a long way at low inference cost—checking for hallucinations, tool usage correctness, and adherence to the knowledge base."

Hillclimbing (self-improvement via hill climbing): In the context of agent engineering, Hillclimbing typically refers to a class of local-search-style automated optimization loops—the system collects failure samples from execution trajectories (such as incorrect tool call parameters or unnecessary hallucinated outputs), automatically generates candidate fixes, and retains only changes that perform better on a validation set, iterating from there. Its advantage is that it requires no gradients or human labeling, directly making local adjustments to prompts, tool signatures, or planning templates. Its limitations are equally clear: a small search radius makes it prone to local optima, and it relies heavily on automatically verifiable evaluation signals (pass/fail, tool return codes). When failure causes require semantic judgment spanning multiple pages (for example, identifying whether a redirect failed due to an expired login session), manual supervision remains irreplaceable.

The Engineer's Perspective: Optimism and Concerns About Skill Atrophy

Shah's career path spans software engineering roles at DoorDash and Adobe, scaling a code data business from zero to eight-figure revenue at Merkle, and co-founding the evaluation platform Pipelines.

Despite building autonomous systems that handle complex enterprise workflows, Shah maintains a grounded perspective on the AI revolution.

"Overall I'm optimistic, but I often think about skill atrophy," Shah admitted. "If I use AI to write code today, and AI to write everything else tomorrow, you risk losing core skills. A good AI engineer first needs a solid engineering foundation."

The core takeaway from this case may be this: while pursuing agent autonomy, defining clear evaluation benchmarks, choosing the right model tier, and identifying which parts can be optimized through hardcoding remain irreplaceable exercises of human engineering judgment.

Skill atrophy: This refers to the decline in an engineer's sensitivity to underlying principles and debugging intuition after long-term reliance on AI for coding, debugging, architecture design, and other tasks. This phenomenon is aligned with the classic concept of "automation complacency" in the automation field—when a system is correct most of the time, human supervisors unconsciously reduce their review frequency until a serious failure occurs. For enterprises, the risk is not AI replacing engineers, but the emergence of a vacuum where "no one can diagnose why the AI went wrong": the model can generate code, but no one can read the subtle security or performance issues within it. Shah's point is therefore worth heeding—a solid engineering foundation is not something to be replaced, but the prerequisite for effectively harnessing AI.

One-line Summary

Brainbase used nine iterations to raise browser agent accuracy from 41% to 91%, cutting costs to $20–30 per week.

Paragraph Summary

Brainbase Labs deployed autonomous browser agents for Domino's Pizza, replacing an external consulting team that manually tracked competitor pricing and promotions. After nine iterations, accuracy rose from 41% to 91%, costs dropped from thousands of dollars per month in consulting fees to $20–30 per week, with inference cost per data source at just $1.16. Engineering lead Milind Shah's key methodology included: co-creating an internal benchmark dataset with the business team before writing code; hardcoding fixed X/Y coordinates to replace visual reasoning for the biggest gains; using highly structured prompts with Sonnet 5 to beat flagship models on both accuracy and cost; and auto-fixing tool-call and hallucination failures via a Hillclimbing engine. Browser navigation trajectories still require manual engineer review, and Shah also flags the risk of skill atrophy.