In an earlier post we made the case that the choice of agent architecture quietly determines whether enterprise AI projects succeed. For the largest class of enterprise work, the recurring, structured, cross-system workflows owned by line-of-business teams, our eight-question framework converged on one architecture: the compiled workflow agent.
Our whitepaper, Enterprise Agents, Compiled, describes that architecture as we have implemented it. This post covers the main ideas.
The cost of intelligence: every loop turn, or once up front
Nearly every agent in production today is built on the agentic loop pattern: an LLM using tools in a loop, deciding each next step by inference until it judges the work done. The loop is the right engine for open-ended work. But every cycle is a billed LLM call, and the plan exists only in the model's context, re-derived from scratch on every run.
Most enterprise workflows are not open-ended. A renewal motion across CRM and the service desk, a vendor-spend motion across procurement and ERP: the steps are knowable and the job repeats. Re-deriving a plan that never changes injects variability and cost into work that should produce the same result every time.
A compiled workflow agent applies the model's intelligence at two stages instead. At authoring time, the model generates and validates the code implementing each step, once. At runtime, the workflow executes as ordinary, deterministic code, invoking the model only where judgment is genuinely required. The LLM is the compiler, not the runtime. Independent research on compiled AI measured 100 percent output reproducibility for this kind of execution, against roughly 95 percent when the model reasons at runtime.
The workspace makes it work
Real workflows bring exceptions and change that deterministic code alone cannot handle. What makes a compiled workflow agent possible is its workspace: a live, two-way copy of everything the workflow touches, continuously synced with the connected systems. It serves three roles at once: it holds the workflow's state, supplies fresh, pre-joined context for every inference, and serves as the agent's memory across runs.
The team works from a live view into that workspace: the data in the systems of record, in one place. Headless SaaS, not five vendor UIs. Questions go to Ask Chat and come back answered from live records, as text, charts, tables, and forms the team can act on in place. Consequential changes wait at an approval gate for a person to approve, adjust, or decline, with every decision recorded and written under the reviewer's identity.
Between decisions, the agent runs unattended on its triggers: schedules, record changes, external events. Salesforce's own research benchmarks leading LLM agents at roughly 58 percent single-turn success on CRM tasks, workable for an assistant with a human watching but not for an unattended hourly job that runs 730 times a month.
Security by construction
A compiled agent's security is a property of how it is built. Everything the agent can do is expressed in a language of canonical nouns (the records and fields, with one name whichever system holds them) and normalized verbs (the common set of operations on them), baked into its code at authoring time. At runtime, every action flows through a gateway under the acting user's SSO identity, so the permissions the enterprise already maintains decide what the agent can see and change. A request outside that language cannot be expressed, let alone executed.
The same construction eliminates prompt injection, the risk that instructions hidden in an email or a ticket comment redirect a loop-based agent mid-run with the agent's full permissions. In a compiled workflow, the steps are fixed code that data cannot rewrite. And every run can be replayed exactly: when compliance asks why the agent did something, the answer is a reproduction, not a reconstruction. How the platform enforces all of this, twice, is covered in the whitepaper.
Intelligence purchased once
We measured the economics directly, building the same resume-screening workflow twice, once using runtime skills and once compiled, and running both against the same fifteen applicants. The skills agent reloaded its instructions, tool schemas, and state on every run, every step a billed round trip. The compiled agent spent tokens in exactly two places: under a dollar of one-time code generation at authoring, and one scoped scoring call per candidate.
The compiled agent cost roughly 11 times less per run. Costs stay flat at any schedule and fleet size, so the economics that justify a pilot still hold for a fleet.
Who can build one
Prompt One's compiled workflow agents add a critical capability: self-serve design and deployment. The line-of-business teams that run these motions (sales, finance, HR, customer success) almost never have engineering resources to automate their own work. With Prompt One, the team member who owns the work describes it in plain language, and the Composer turns the description into a deployed agent, generating code using only the data and operations the author is entitled to. Deployment is one click, teammates subscribe to a finished workflow in one click, and no step waits on an engineering queue.
The whitepaper covers all of this in more detail, along with the full agent lifecycle: how a plain-language description becomes a Motion, the shareable definition of a team's recurring work, and how it becomes something the whole team runs. It is a short read: download it here. And if your team has a recurring motion it wants automated, you can start a free trial and compile it yourself.