AaaS
How it works
The client defines a goal or uploads source materials (procedures, call transcripts, recordings, documentation). The provider's agent (e.g., Ghostwriter) analyzes this data, identifies key behaviors and edge cases, generates a production-ready executive agent, and configures it across multiple channels (voice, chat, email) with built-in safeguards. A continuous improvement loop then analyzes real interactions, proposes enhancements, tests them in a sandbox, and prepares them for human approval. All interaction with the agent occurs in natural language โ no UI clicks required.
Problem solved
The traditional SaaS model requires a human to learn an application's interface and perform work step by step through manual interaction. This scales linearly with the number of users and their cognitive constraints. AaaS addresses this by delegating execution to an AI agent, reducing human interaction to defining the goal and accepting the result.
Components
Layer that provides the agent with tools, memory, planning, a coherent action space, and the right task context. Without it, the agent has no stable foundation for production operation.
Refactoring of the SaaS platform so that all functions are accessible programmatically (API, SDK) without UI dependency. Lets the agent invoke the platform directly instead of emulating clicks.
Isolated space where the agent builds and tests changes before deploying to production. Critical for safe autonomy โ the agent can experiment without risking damage to the running system.
Official
Cycle of analyzing real interactions, identifying improvement opportunities, validating them, and preparing them for review. Runs autonomously in the background and lets agents improve over time.
Official
Gate where changes prepared by the agent are approved by a human before deployment. Forms the foundation of trust and accountability in production AaaS deployments.
Official
Implementation
Attempts to implement AaaS on a UI-first platform (clicks, forms) result in an agent emulating a human user โ unstable and slow. Without refactoring the platform into headless infrastructure, the paradigm does not work.
Using a seat-based billing model for AaaS weakens the value proposition โ the customer doesn't buy seats, they buy outcomes. Misaligned billing obscures ROI and slows adoption.
Full agent autonomy from day one is risky โ without an observed phase and incremental permission delegation, the customer won't build trust, and agent errors can undermine the whole contract.
AaaS relies on continuous improvement. Without automated evaluation of real interactions, the agent assembly line has no signal โ the agent stops learning and drifts away from business changes.
For deterministic, well-defined tasks (forms, simple CRUD), AaaS adds LLM cost, latency, and unpredictability. Traditional SaaS is often faster, cheaper, and more predictable.
Evolution
Salesforce introduces the SaaS model based on a centrally hosted, human-operated web application. It defines the 'customer buys access to an interface' paradigm against which AaaS will later position itself.
Yao et al. (2022) show that LLMs can act as a reasoning engine in loops combining thoughts with tool actions. This is the technical substrate on which AaaS becomes possible.
Anthropic publishes guidelines distinguishing workflows from agents and formalizing five composition patterns. Practitioners gain a common vocabulary that will later ease AaaS commercialization.
On March 25, 2026, Bret Taylor and Clay Bavor (Sierra co-founders) publish the Agents as a Service manifesto and introduce Ghostwriter โ an agent that builds agents. They coin the phrase 'prompts, not clicks' and the 'agent assembly line' concept. This is the moment the term enters public industry discourse.
Technical details
Hyperparameters (configurable axes)
Scope of decisions the agent can make without human approval โ from proposals requiring review to full autonomous deployments.
Billing model: per seat (SaaS-like), per usage (per token / call), or per outcome (resolved case, closed ticket).
Number and type of channels the agent operates on: chat, voice, email, video, messengers.
Format in which the customer defines what they expect from the agent: text prompt, process documentation, transcript, audio recording, whiteboard photo.
Level of built-in safeguards constraining agent actions: content filtering, compliance verification, tool validation.
Execution paradigm
AaaS is a delivery paradigm, not a specific computational kernel. The execution mode is inherited from the underlying Agentic AI: conditional loops driven by an LLM over the platform's headless infrastructure.
The vendor's agent decides which platform components to invoke, what kind of production agent to create, and when to request human approval. Routing is not predefined โ it follows from analysis of the customer's goal and available tools.
Parallelism
Parallelism occurs primarily across clients (different clients, different production agents) and within multi-agent subsystems (e.g., concurrent sandbox tests).
Hardware requirements
AaaS is a delivery model, not a computational kernel. Hardware requirements stem from the underlying LLMs and platform tools, not from the service paradigm itself.