SecondBrain
Ask the Brain
Index/Sourceupdated Sat May 30 2026 08:00:00 GMT+0800 (Philippine Standard Time)

You're Not Behind (Yet) Learn AI Agents (theMITmonk)

agentsagentic-ainarrow-agentsoodaai-literacybrand-fodder

You're Not Behind (Yet) Learn AI Agents (theMITmonk)

Fourth theMITmonk video (Sandeep Swadia) in this wiki. A 13-minute primer on what makes an agent an agent (vs a prompt or a workflow), how to decide what to automate, and why narrow agents win. Brings three new named frameworks to this vault: ARR Framework (prompt-vs-agent test), OODA Loop (adapted from John Boyd), and GPS Check (for Agents) (Goal / Proof / Steps). Closes on the narrow-ownership thesis that pairs with Build vs Buy (Agents) and Bounded vs Unbounded Tasks.

The core distinction

"A chatbot waits for your next prompt. An agent figures out its next move."

Sandeep's pithiest framing: a chatbot predicts the next word; an agent decides the next action.

Driver analogy: a prompt is sitting next to a student driver (you stay alert, you correct). An agent is a hired driver — "You set the destination, hand over the keys, and just sit in the back seat. It handles the route, the traffic, and all the step-by-step decisions."

ARR Framework — when to use an agent vs a prompt

"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."

See ARR Framework. Cleanest single-axis decision rule for "should this be an agent?" in the vault.

Under the hood — the four roles

Sandeep's mental model of what happens inside an agent (one LLM at the center, four workers around it):

  1. Analyst — finds the pattern in raw inputs
  2. Planner — decides what matters and what to do
  3. Operator — does the work / makes calls / produces output
  4. Auditor — checks the result, refines

Worked example: "Every Monday at 7am: review last week's customer support tickets, sales notes, and product feedback. Identify three biggest recurring issues, summarize what changed, and email my leadership team a one-page weekly brief."

This pairs with the existing Agentic Loop / ReAct framing in this vault. Same underlying mechanism (reason → act → observe → repeat); Sandeep's framing assigns role labels to each pass instead of step labels — useful for explaining to non-technical audiences.

OODA Loop — adaptation as the real differentiator

"The best thing about agents is that they can adapt when things go wrong."

Anchor story: John Boyd's 1970s research into why American F-86 pilots beat technically superior Soviet MiG-15s in Korea. Conclusion: "The American pilots could see more from their cockpits, and they could adapt faster." They got inside the enemy's decision cycle. Boyd called this OODA: Observe, Orient, Decide, Act.

Sandeep transplants OODA to agents: the test of an agent is not whether it can run the obvious path; it's whether it can re-plan when that path breaks.

"A workflow can follow the process. An agent can reroute it completely."

Worked example: a Friday grocery-shopping agent's usual item is out of stock + six friends coming over Saturday. The workflow breaks. The agent: sees the gap, finds substitutes, adjusts quantities for six, checks the calendar, rebuilds the order.

See OODA Loop for the dedicated page.

Why agents fail — the GPS Check (for Agents)

"The most dangerous thing about AI agents is that they will do wrong things faster and with more confidence than you ever could. An agent is not magic. It's a multiplier."

CMO anecdote: "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." Sandeep's frame: "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."

The diagnostic before automating anything — GPS:

  • Goal — "Can I define the goal in one sentence very clearly?"
  • Proof — "Can I tell what good looks like? And how will I know if the agent got it right?"
  • Steps — "Can I describe each and every step very clearly without a lot of handwaving?"

Worked contrast:

  • Weak: "Summarize my emails every morning."
  • Strong: "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."

"That gap is exactly where the mess lives."

See GPS Check (for Agents).

The narrow-ownership thesis

"Most companies want AI everywhere. The ones actually winning are obsessively narrow. If clarity is the bottleneck, then the opportunity is not broad intelligence, it's narrow ownership."

Anchor story: construction-software customer conference, product lead demos a single agent solving one specific field-data-collection workflow. "The demo worked mostly with a few minor glitches here and there, but when he showed the QR code at the end, every hand in the conference room went up with their phones."

Sandeep's test: "Find a highly specific task people hate doing, but they have to do it repeatedly. That's where the money is."

"For every software company that exists today, there will be an agent company trying to dethrone it. The winners won't build the broadest agents first. They'll build the one that understands one workflow, one market, and one kind of user pain better than everyone else."

Pairs directly with Build vs Buy (Agents) (Praveen's front-end-buy / back-end-build split) and Bounded vs Unbounded Tasks (narrow = bounded). Sandeep's contribution is the go-to-market sharpness — the founder/operator framing where the existing pages are enterprise-architecture framings.

Closing thesis — output infinite, judgment scarce

"We're entering an era of infinite output, content, code, and analysis all becoming super cheap. When intelligence becomes that cheap, judgment becomes even more expensive. When output becomes infinite, taste becomes scarce."

"The most valuable person is no longer the one who can think the fastest. It's the one who can define good work, spot bad work, and know when to trust an agent and when to trust a human."

Directly aligns with Code Is Free (Lopopolo) and the floor-vs-ceiling motif from Vibe Coding / Agentic Engineering. Sandeep's contribution: explicitly names judgment and taste as the new scarce resources (where Lopopolo focused on tokens/attention and Karpathy on understanding). All three are talking about the same shift.

Why this matters to the vault

  • First ARR taxonomy in the vault. Augments Bounded vs Unbounded Tasks (architecture-side) and DRAG Framework (personal-productivity-side) with a single one-line decision rule. The three frames compose: ARR is the quick triage; DRAG is the delegation rubric for the prompt half; Bounded/Unbounded is the autonomy-readiness side for the agent half.
  • OODA gives the vault a vocabulary for agent adaptation it didn't have. Agentic Loop covers the mechanism; OODA covers the adversarial speed-of-decision dimension. Useful for thinking about real-time agent vs human comparisons.
  • GPS Check is the action-side precondition for AWARE Framework — before deploying with technical controls, can you even articulate the goal/proof/steps? GPS is the cheap pre-flight; AWARE is the runtime/governance layer.
  • Narrow-ownership thesis hardens the Build vs Buy (Agents) page with a go-to-market lens.
  • Judgment-scarce closing pairs with Code Is Free for a fifth source on the same value-relocation argument.

Editorial pattern (4th theMITmonk source — pattern is solid)

  • Solo monologue, 13 min
  • Named acronym frameworks (ARR + GPS in one video — Sandeep's framework-density is the highest in the vault)
  • Anchor story + cognitive-science / military-history citation (OODA + Boyd here; Ebbinghaus + Bjork previously)
  • Sponsor: newsletter only — no commercial sponsor
  • Personal anecdotes are board-room / advising-CMO flavored — confirms the "advising billion-dollar companies" persona

Practical takeaways for this vault's user

  • Brand fodder candidate — "Most AI problems are human problems in disguise" + the CMO "right people in the seats first" line is a near-ready Medium / LinkedIn post on AI deployment failures. On-thesis for the senior-IT-leader audience because it reframes AI strategy as organizational readiness, not tool selection. Pairs cleanly with the existing CIO Agenda 2026 (CXOTalk) "88% use AI, <6% get value" statistic.
  • GPS Check is directly usable as a coaching artifact. Three-question pre-flight for any team-member proposing an agent. Lighter-weight than AWARE Framework for early-stage triage.
  • OODA / narrow-ownership pair is a thought-leadership essay candidate — "the agents that win aren't the broad ones, they're the ones that can OODA-loop inside a workflow your competitors are still hand-rolling."

Cross-links

Source

  • Original transcript