Not live yet

What is the Agentic Workload Compiler?

You describe your agentic network — agents, tasks, workflows, tools and business volume. The compiler derives the inference workload that network creates, instead of making you guess token counts. It then runs that workload against latency vs batching scenarios to produce the report you actually want: how many tokens per month, which model harness, and a defensible cost figure to quote your client.

The compiler boundary works right now — no signup needed to try it. The report engine lands only if developers signal they want it. That's what the waitlist is for.

  • Early access to the report engine when it ships
  • Early-bird pricing for the first cohort
  • A say in what gets built next
Functional prototypeCompiler boundary only

Agentic Workload Compiler

Turn an agentic solution into a standardised workload model — without asking the developer to estimate token usage.

Describe your agentic network once. The compiler derives what inference workload it creates — then price that against the latency vs batching tradeoffs your client actually cares about.

Developers provide the solution

Agents, tasks, workflows, tools and business volume. No token counts.

The compiler determines the workload

Inference operations, execution profiles, token profiles and batching opportunities — derived, not requested.

How it works

Step 1

Describe

Agents, tasks, workflows, tools and business volume. No token counts.

Step 2

Compile

Derives the workload — inference operations, token profiles, execution opportunities.

Step 3

Simulate

Runs the dataset across scenarios: all-realtime, mixed batching, all-batching.

Step 4

Report

Tokens / month per scenario, recommended harness, and a $ / month figure to quote.

The report you take to a client: “running this network this way will use roughly X tokens a month — under all-realtime, mixed-batching or all-batching. Here's the recommended model harness and what it will cost them.” No more estimating token usage on a whiteboard.

This prototype demonstrates steps 1–2 and the simulation seam (the workload dataset). The report engine, provider pricing and harness recommendation land after market validation — the waitlist is the signal that decides it.

Stage 1 of 5

Solution

Describe what the system does: agents, tasks, workflows and tools. You describe the solution; the compiler derives the workload from its structure.

Agents

Each agent carries a purpose and a model role. The compiler uses these to derive inference operations — you never describe tokens.

Tasks

Tasks carry the business volume. Volume is developer input — the compiler derives operations-per-work-unit and total inference operations from the workflow.

Workflows

The routing between agents. Dependencies are resolved from these links.

  • OrchestratorKnowledge Agent
  • OrchestratorCRM Agent
  • Knowledge AgentResponse Agent
  • CRM AgentResponse Agent

Tools & Models

Tool operations are distinct from inference operations — the compiler keeps them separate.

Knowledge Baseknowledge retrieval

Retrieves relevant product/support documentation for a query.

CRMdatabase

Looks up customer, account and order information.

Ticketingdatabase

Creates or updates support tickets.

Webhookhttp

Posts enrichment events downstream.

Model ARouting and intent classification
Model BKnowledge-grounded reasoning
Model CStructured data extraction / tool orchestration
Model DFinal customer-facing generation

Support the build

The compiler boundary is live today. The report engine is next — and it ships only if the market signals it. Join the waitlist for early access and early-bird pricing when it does.

Report a problem