An independent study reference written by Dr Phuc V. Nguyen. It is not official subject material — for assessment requirements always follow your subject outline and vUWS.

The Analyst and What Comes Nextintermediate

Agentic AI

Q: An agent completes a nine-step reporting task. Each step is right about 90% of the time given that the steps before it were right, and every step must be right for the report to be right. Roughly how often is the finished report correct, and what follows from that?

About 39%, so the effective fix is checkpoints inside the chain rather than a marginally better model. Multiply the conditional step reliabilities: 0.9 to the ninth power is about 0.39. Per-step accuracy that sounds fine collapses across a chain. Lifting every step from 90% to 95% only raises the end result to about 63%, so the remedy is to break the chain with verification points where an error can be caught before it propagates.

Q: Two proposals for the same agent. Proposal A lets it run any query but requires human approval before it writes to any table. Proposal B lets it write freely but caps it at fifty actions a day. Which is the better default, and why?

A, because it separates actions by reversibility, whereas a count limit still permits an irreversible write. A cap on volume does not distinguish a harmless query from a table overwrite. Fifty actions are more than enough to do lasting damage if one of them is destructive. Gating on reversibility lets the cheap, recoverable work run unattended while the consequential work waits for a person, which is where the risk actually sits.

An agent is a system that pursues a goal with some autonomy. It decides for itself what to do next, reacts to what it observes and keeps going without being told each step. The classical definition rests on autonomy, reactivity and pro-activeness rather than on any one technology. The pattern now common in analytics is a tool-using one: a model runs a loop of observe, decide, act, observe again, calling out to warehouses or ticketing systems and changing something outside itself. Agents are being pointed at pipelines, query writing, monitoring and reporting. The engineering question is how far that loop may run before a person checks it.

Why it matters

Compare two assistants. The first answers questions and stops. You ask for last quarter's churn, you get a number, and everything after that is yours to do. The second is given the goal, opens the warehouse, writes the query, notices the region codes changed in June, patches the join and emails the result. The second is far more useful, and far more capable of being confidently wrong at four in the morning with nobody watching.

Before you read on — recall

An agent completes a nine-step reporting task. Each step is right about 90% of the time given that the steps before it were right, and every step must be right for the report to be right. Roughly how often is the finished report correct, and what follows from that?

Formulas

Reliability across a chain of steps

P(\text{all } n \text{ steps succeed}) = \prod_{i=1}^{n} p_i = p^{\,n}

Let

p_i

be the probability that step

i

succeeds given that every earlier step succeeded. Defined that way the product is exact, and the last equality holds only when that conditional probability is the same

p

at every step. At

p = 0.95

, a five-step task finishes correctly about 77% of the time and a twenty-step task about 36%. Two cautions. Accuracy measured on a step in isolation is not

p_i

, and agent errors can be correlated, because one wrong assumption early can spoil several later steps at once. A naive product of isolated per-step accuracies can therefore mislead in either direction. The lesson survives the caveats: per-step reliability that sounds acceptable becomes unacceptable end to end, which is why checkpoints belong inside the loop rather than only at the end of it.

Worked examples

Scenario

A finance team gives an agent a standing goal: each morning refresh the revenue dashboard, investigate any region that moved more than 10%, and post a short note. After twenty-one days it posts a note saying revenue in one region has collapsed.

Solution

Trace the loop before believing the note. The agent queried, compared, found a drop and wrote it up, and every one of those steps was individually correct. The cause sat upstream: a source system began sending that region under a new code mid-month, so half its sales landed in an unmatched bucket. The agent had no way to know the definition of the region had changed. Agents inherit whatever data quality sits beneath them, and a fluent explanation of a broken number is more dangerous than no explanation at all.

Scenario

An analytics lead is asked to let an agent open and close incident tickets on the production pipeline without approval.

Solution

Split the actions by whether they are reversible and how far the consequence travels. Reading logs, running a diagnostic query and drafting a ticket are cheap to undo, so let the loop run. Restarting a job, backfilling a table or closing an incident changes state that other people depend on, so those take an approval step. Bound the loop as well: a ceiling on actions per run, a whitelist of tools it may call, and a stored record of every call and every value returned so the work can be audited afterwards. Autonomy is granted per action, not per system.

Common mistakes

✗An agent is a chatbot with a better prompt. The defining feature is not the wording of the request but the loop: the system decides what to do next, observes the consequence and goes again. Acting on outside systems through tools is the usual shape of that loop in analytics work, rather than a test of what does and does not count as an agent.
✗More autonomy makes an agent more useful. Usefulness peaks well short of full autonomy for most business tasks, because the cost of a wrong action is rarely symmetric with the benefit of a right one. The design question is which actions are reversible, not how independent the system can be made.
✗If each step is reliable, the whole task will be reliable. Step reliabilities compound. A chain of twenty steps, each right 95% of the time given that the steps before it were right, finishes correctly about a third of the time, so long chains need verification points rather than marginally better single steps. Correlated errors make the arithmetic harder still, because one wrong assumption can spoil several steps at once.
✗An agent that explains its reasoning has shown you its reasoning. The explanation is generated text about the work, not a trace of the work. The audit trail is the log of tools called and values returned, and that is the artefact worth keeping.

Revision bullets

•Agent = a goal plus a loop of observe, decide, act, observe
•Unlike a plain model, the common analytics pattern acts through tools
•Step reliabilities compound along a chain, so long chains degrade quickly
•Grant autonomy per action, graded by how reversible the action is
•The audit trail is the tool-call log, not the generated explanation
•Agents inherit the data quality of everything beneath them

Quick check

Two proposals for the same agent. Proposal A lets it run any query but requires human approval before it writes to any table. Proposal B lets it write freely but caps it at fifty actions a day. Which is the better default, and why?

Connected topics

Decision Support NLP Into Production Algorithmic Bias Human-AI Agency

More in The Analyst and What Comes Next

Analytics as a Service Analytics Roles Changing Analyst Role

Sources

Wooldridge & Jennings (1995)
Wooldridge, M., & Jennings, N. R. "Intelligent Agents: Theory and Practice." The Knowledge Engineering Review, 10(2), 115-152, 1995.
The durable definition of an agent through autonomy, reactivity, pro-activeness and social ability, written long before the current wave of tools.
Russell & Norvig (2021)
Russell, S., & Norvig, P. Artificial Intelligence: A Modern Approach. 4th ed. Pearson, 2021.
Frames the whole field around the rational agent and the perceive-act cycle, which is the loop described in this node.

How to cite this page

Dr. Phil's Quant Lab. (2026). Agentic AI. Derivatives Atlas. https://phucnguyenvan.com/concept/ba-agentic-ai