ARR Framework
ARR Framework
Sandeep Swadia's one-axis decision rule for whether to use an agent or a prompt, from You're Not Behind (Yet) Learn AI Agents (theMITmonk).
"If a task is autonomous, recurring, and reviewable, it's a strong candidate for an agent. If it needs live judgment or it only happens once or can't be reviewed clearly, then use a prompt."
The three tests
| Letter | Test | Why it matters |
|---|---|---|
| A | Autonomous | Can it run without you needing to make in-the-loop calls? If every step needs your judgment, you're not building an agent — you're using a prompt. |
| R | Recurring | Does it happen often enough that automating it pays back? One-shot tasks belong in the prompt column even if they're autonomous. |
| R | Reviewable | Can you tell whether it did a good job? If "good" is irreducibly subjective or invisible to the user, the agent has no feedback signal and will drift. |
Worked contrast
- Prompt: "Write me a LinkedIn post." (One-off, judgment-heavy)
- Agent: "Watch my industry every Monday. Find the three most relevant stories. Study my previous posts. Draft a new post based on those stories in my voice. Revise against my style and schedule it for Tuesday morning."
The agent version satisfies all three tests; the prompt version fails at least Recurring (and arguably Reviewable for "what makes a LinkedIn post good?").
How it sits among other delegation frames in this vault
| Frame | Question it answers | Author |
|---|---|---|
| ARR | "Should this be an agent or a prompt?" | Sandeep |
| DRAG Framework | "What kind of work should I delegate to AI at all?" | Sandeep |
| Bounded vs Unbounded Tasks | "Is this task safe to make autonomous in the enterprise?" | Praveen Akkiraju |
| Build vs Buy (Agents) | "Should we build or buy this agent?" | Praveen Akkiraju |
ARR is the fastest screen. If ARR says "prompt," the other questions don't matter for this task. If ARR says "agent," then Bounded vs Unbounded Tasks tells you how aggressive to be with autonomy, and Build vs Buy (Agents) tells you whether to roll your own.
Where ARR is incomplete
The Reviewable test is doing a lot of work — it conflates can a human review with can the agent self-evaluate. The richer treatment in this vault is on the AWARE Framework page (the "ecosystem observability" pillar) plus LLM as Judge (auto-evaluation pattern). ARR is the bumper-sticker; AWARE is the detailed checklist.