GPS Check (for Agents)
GPS Check (for Agents)
Sandeep Swadia's pre-flight diagnostic before automating anything with an agent: Goal / Proof / Steps. From You're Not Behind (Yet) Learn AI Agents (theMITmonk).
"An agent is not magic. It's a multiplier. ... An agent doesn't fix bad thinking. It formalizes it. Usually the agent fails because the human was vague, not because the underlying model was bad or anything."
The three questions
| Letter | Question | Failure if you can't answer |
|---|---|---|
| G | Goal — "Can I define the goal in one sentence very clearly?" | The agent optimizes the wrong objective; you don't notice for a while |
| P | Proof — "Can I tell what good looks like? And how will I know if the agent got it right?" | No feedback signal; agent drifts; you can't audit it |
| S | Steps — "Can I describe each and every step very clearly without a lot of handwaving?" | Agent reasons the steps itself, often wrong, often confidently |
"Unless you can do those three things very well, your agent is not going to make any difference."
Worked contrast
| Weak | Strong |
|---|---|
| "Summarize my emails every morning." | "Every morning at 7am, read my unread emails, categorize them by urgency, draft replies to routine messages, and flag anything from my top five customers." |
The gap between those two prompts is exactly the GPS work the human is supposed to do before deploying.
Sandeep's diagnosis of why agents fail
"Most AI problems are human problems in disguise. An agent is just a mirror. It reflects the quality of your thinking back at you. It just amplifies it. Give an agent vague goals, sloppy directions, and no way to get feedback, and it will drive the car straight into the tree faster and with more confidence than you ever could."
The CMO anecdote in the source: a profitable consumer company asked Sandeep what was stopping them from using AI for customer acquisition. The CMO: "We have all the data, but we'll still need to build a clean process so we can turn that into something useful... we need the right people in the seats first."
GPS is the pre-condition for the agent to add value at all.
Where it fits in the vault
- Pre-condition for ARR Framework — even an autonomous/recurring/reviewable task fails GPS if the goal is fuzzy. ARR says "can this be an agent?"; GPS says "are you ready to deploy an agent on this?".
- Cheaper triage than AWARE Framework — AWARE is the runtime / governance control structure; GPS is the 3-minute sanity check you do before even getting to the governance review.
- Echoes the CIO Agenda 2026 (CXOTalk) / Token Maxing failure pattern: orgs blow through AI budgets without GPS in place, then call AI strategies a failure. The Crawford/Sacolick "88% use AI, <6% get value" data point is the org-level shadow of GPS failure at the individual-deployment level.
- Connects to Productive Resistance / Intelligent Hill (Prompting Camps) — the Interview step of PRIME Framework is what AI does for you when you skipped GPS; chain-of-thought prompts force the LLM to surface the steps; both are remediation. GPS is the prevention.
The narrow-ownership corollary
Sandeep ties GPS to the Narrow Agents thesis: "The winners who can wield the power of AI agents aren't just going to be engineers. They'll be the people who understand their work deeply enough to define it precisely." Domain ownership = ability to pass GPS = competitive edge.
Practical takeaway for this vault's user
GPS is small enough to use as a 3-question gate for any team member proposing an agent build. "Tell me your G, P, and S in three sentences — if you can't, the agent isn't ready." Functions as an org-level filter against the bad-thinking-amplifier failure mode without requiring the heavier AWARE Framework review.