MIT documented why. Systems do not retain feedback. Do not adapt to context. Do not improve over time. The industry response has been to scale model capability — bigger context windows, more parameters, frontier prices. A more capable model still drifts, still produces false confidence, still operates without context. Capability does not address architectural failure.
Decompose the cognitive work into specialised blocks. Route each block to the cheapest model that can do its job reliably. Run an independent evaluator that catches failures before they ship. The architecture does what scaling capability cannot.
Same document, same task, three price points. The frontier prompt approach on the left. The Agxel engine on a workhorse model in the middle. The Agxel engine on a budget open-source model on the right.
The standard way to summarise a trial publication: drop the PDF into a frontier model with one large prompt. Most enterprise AI deployments work this way.
The Agxel engine works differently. The task decomposes into a pipeline — extract, validate, interpret, write, package. Each step runs on the cheapest model that can do that specific job reliably. An evaluator runs across all of them, catching drift before output.
Below: the same trial PDF processed three ways in parallel.
Works best with peer-reviewed trial publications and technical reports. Results vary with document complexity.
Drop a PDF here or click to browse
Max 10 MB
Runs typically take 5–10 minutes. All three models process in parallel — go make a coffee.
Same input. Same task. Three radically different cost profiles.
| Approach | Model | Wall time | Input tokens | Output tokens | Cost |
|---|
After reading the three outputs:
How much editing would each of these need before you'd send it to a customer?
This is one engine. The same architecture runs five others, each built for a specific cognitive task in commercial workflows.
If you want to see what your team's specific workflow could look like:
A prompt is a single instruction sent to an AI model. The simplest unit. The model receives the prompt plus any attached content (a PDF, an image), produces one response, and that's it. No memory of previous calls, no tools, no follow-up. The "Prompt" column on this page is exactly this: one large instruction, one model call, one output.
An agent is an AI model that can call tools — search the web, read files, run code, query a database — and use the results to make its next decision. An agent loops: think, act, observe, repeat, until the task is done. Most consumer "AI assistants" are agents.
A skill is a packaged capability that an agent can load on demand. A specialised instruction set with optional reference files — for example, a "write a Linear ticket" skill or a "summarise a meeting" skill. Agents invoke skills when the task at hand calls for one.
An Agxel engine is none of these on its own — and uses all of them. An engine is a structured pipeline of specialised steps, each running on the cheapest AI model that can do its specific job reliably. Some steps are simple prompts (extract data from a PDF). Some steps are small agents (decide what type of trial this is). The pipeline is orchestrated: a separate orchestrator decides which step fires when, with quality gates between them that can reject and re-run a step. Unlike a single agent, the engine is deterministic about sequence and order. Unlike a single prompt, it can validate and repair its own intermediate outputs. The result is more reliable, more auditable, and dramatically cheaper than asking one big model to do everything in one go.
Because the engine doesn't need a Frontier model for every step. Extracting structured data from a PDF doesn't require Frontier-level intelligence. Writing a one-pager from already-extracted facts doesn't either. Only one or two steps in the pipeline — the interpretation and writing steps — need more capability, and even those run comfortably on a Standard model. The Budget version of the engine pushes the same logic further, running every step on a low-cost open-source model.
You pay for the intelligence you actually need at each step, not the intelligence the hardest step required.
Frontier is the most capable model in the family — Claude Opus 4.7 or GPT-5.5. The most expensive (Opus runs over $1 per shot, GPT-5.5 around $0.35), the smartest, and on this page actually the fastest in wall time — because the Prompt column is just one API call, where each Engine column is six steps in sequence.
Standard is the workhorse model — Claude Sonnet 4.6 or GPT-4.1. Roughly 5–10× cheaper than the frontier, similar per-token speed, capable enough for most production tasks.
Budget is an open-source model running on third-party infrastructure — currently Nvidia's Nemotron Super 49B. Roughly 100× cheaper than the frontier model and genuinely slower per token. Surprisingly capable inside a well-designed pipeline, but the longest wall time on this page.
Speed on this page is dominated by pipeline depth, not by model tier: a single frontier-model prompt finishes faster than any six-step engine, even one running on the same family.
Different methodology, different output shape. The prompt produces what the model decides to produce. The engine produces what each step is designed to produce — which means more consistent structure, exact numbers from extracted tables, and a quality gate that rejects thin output.
You will often see the engine output is more thorough (because the extraction step captures everything before any summarisation begins) and more honest about gaps (because validation flags missing data explicitly).
Real model calls. The prompt is one big API call against a frontier model on a 50+ page PDF. The engine is six API calls (extract → validate → interpret → write → potentially re-write → one-pager). All three run in parallel, but the slowest one sets the wall time. The budget engine is the slowest because the budget model is genuinely slower per token.
This is honest end-to-end production time. Nothing cached, nothing pre-computed.
Yes — use the "Upload your own" tab and drop in any trial publication or technical PDF (up to 10 MB). The same pipeline runs against your document. Nothing is stored beyond the run.
If you want to discuss using engines like this on your own product portfolio, book a 20-minute walkthrough.
MIT's State of AI in Business report found that 95% of enterprise AI pilots fail to deliver measurable results. The dominant cause is not model quality — it is that systems lack feedback loops, context onboarding, and improvement mechanisms. These are architectural problems. Scaling model capability does not solve them. The Agxel engine architecture addresses them directly.
The full text of the prompt sent to the frontier model in "The Prompt" column: