Workflow
Predefined code paths
LLMs and tools orchestrated through paths the system designer wrote. The model fills in the steps; it does not choose them. Predictable, debuggable, cheap, easy to evaluate. Most production AI lives here.
Definition·May 2026·Prism
What is an AI agent?
The industry can’t agree.
We’ll go first.
A common vocabulary, before the buyers ask. One sentence, then the frame, then the primitives, written down so the next conversation can start somewhere instead of nowhere.
I·The Industry Hasn't Agreed
TechCrunch put it bluntly in March 2025: “No one knows what the hell an AI agent is.”
Two years of marketing the term and the people running the largest AI companies on Earth still describe it four different ways. The vocabulary differs; the shape underneath does not.
Sam Altman
OpenAI
“Programs that take actions in the world on your behalf.”
Satya Nadella
Microsoft
“Software coworkers with their own decision-making.”
Marc Benioff
Salesforce
“The next generation of AI that completes tasks end-to-end.”
Dario Amodei
Anthropic
“A model that uses tools in a loop to accomplish a goal.”
II·Our Canonical Definition
One sentence. No mysticism. No special category of software.
An AI agent is a system where a language model decides what to do next, uses tools to act, holds state across steps, and keeps going until the goal is met or a guardrail stops it.
An agent is a program. The novelty is the reasoning loop is delegated to a language model instead of hand-coded branches. Every other property, tool use, memory, autonomy level, planning style, is a knob on that one shape.
Every Prism feature, prompt, and marketing surface is held against this sentence. If something we build contradicts it, we change the build, not the sentence.
III·The Most Useful Frame, Workflow vs Agent
The single most leveraged distinction in the industry, and the field has rallied around it.
From Anthropic’s Schluntz and Zhang (“Building Effective Agents,” December 2024). Two shapes. They look superficially similar; they fail in structurally different ways.
Agent
Dynamic control
LLMs that direct their own processes and tool usage, maintaining control over how they accomplish tasks. Flexible, capable of novel work, harder to debug, harder to bound costs. The right shape for open-ended problems.
Anthropic’s own guidance: “find the simplest solution possible, and only increase complexity when needed.” Prism’s default is workflow; autonomy is earned per surface as the user’s trust accumulates.
IV·The Spectrum
Eight rungs. Three of them are agents; the rest are not.
When we talk about “what kind of agent are we building,” we point at one rung. The vocabulary stops being mushy the moment everyone is pointing at the same ladder.
- 01
LLM call
One prompt. One response. No tools, no memory, no loop.
- 02
Augmented LLM
The model plus retrieval, tools, and memory, but the caller still drives.
- 03
Workflow patterns
Prompt chaining, routing, parallelization. The system designer decides the steps; the model fills them in.
- 04
Orchestrator-Worker
A planner LLM delegates sub-tasks to dynamically chosen worker LLMs.
- 05
Evaluator-Optimizer
A generator LLM and a critic LLM in a feedback loop. Output improves until the critic approves.
- 06
Agent
An autonomous loop where the model picks the next step, uses tools, observes results, and continues until the goal is met or a guardrail stops it. This is where the word agent properly applies.
- 07
Multi-agent system
Multiple agents collaborating across vendors via A2A protocols. Specialization beats single-agent generality.
- 08
Self-improving / lifelong learner
Voyager-style skill library. The agent extends its own capabilities over time. Research frontier.
Rungs 06-08 are properly agents. Everything above is the substrate agents are built on top of.
V·The Six Primitives
Every agent has these six parts. Name them. Map them. Then build.
The industry converged on this list. Prism’s runtime maps to them one-to-one, naming the parts is the trust play.
Model
The reasoning core. Decides what to do next given context.
In Prism: Anthropic Claude, Haiku 4.5 / Sonnet 4.5 / Opus 4.7. Routed by goal type. Never default-locked to one vendor.
Tools
How the agent affects the world. Standardized industry-wide via MCP + function calling.
In Prism: Our tool registry ships as Model Context Protocol servers. Every tool the agent can reach is wire-compatible with Cursor, Claude Desktop, and the rest of the MCP ecosystem.
Memory
What the agent remembers within a run and across runs. Short-term, long-term, semantic.
In Prism: Three layers: raw turns, compressed prose summary, and semantic vector recall. The compression janitor runs nightly so memory grows without inflating context costs.
Planning
How the agent decomposes a goal into steps. ReAct, Reflexion, plan-then-execute, prompt chaining.
In Prism: Implicit in the model's chain-of-thought today; surfaced visually in our Glass Box trace. Workflow mode makes the plan explicit; agent mode lets the model author it.
Reasoning loop
The execution engine that ties Model → Tools → Memory → Planning together.
In Prism: A bounded autonomous loop with hard caps on read and write tools per step. Mixed-trust steps halt the loop and queue the unapproved tools for human review.
Environment + feedback
The world the agent acts in, and the signal it gets back. Human-in-the-loop checkpoints, evaluator models, test runners.
In Prism: Human-in-the-loop approval queue, per-tool trust progression, scheduled wakes, eval-on-every-change, continuous production sampling, and an action receipts log with one-click undo.
VI·Plain-Language Analogies
For the buyers who do not run engineering teams.
An agent is a contractor, not a power tool.
You describe the job; they decide which tools to grab.
A workflow is a recipe; an agent is a chef.
The recipe is steps in order. The chef knows the goal and picks the techniques.
MCP is the USB-C port for AI.
One shape, every device fits.
Memory is a personal assistant's notebook.
What they write down between meetings is what makes them useful tomorrow.
VII·Why We Publish a Definition
MCP became the industry’s tool-integration standard for a reason worth copying.
Anthropic published a spec, shipped working clients, and donated it to the Linux Foundation before the market settled. By late 2025 it was at 97 million monthly SDK downloads and 10,000-plus public servers. The vocabulary won because someone showed up early and wrote it down.
Prism’s leverage is the same posture, one floor up. We do not need to win the definition war. We need to be the place where the answer is already written down, embedded in a product that runs the definition end-to-end, so the buyers carry the language back into their organizations without us having to push it.
We are not arbitrating. We are writing it down.
This page is the canonical reference. If the definition evolves, this page evolves first and the rest of Prism follows.
Build an agent
May 2026 · Prism · Definition · See the Manifesto