Agentic AI Pitfalls: Blame the Harness, Not the Model

Agentic AI pitfalls header for November, month of gratitude and Thanksgiving, with the thesis Blame the Harness, Not the Model, three books on agentic AI, pumpkins and maple leaves, by Dr. Harish Kotadia, Ph.D.

Most agentic AI pitfalls are harness failures, not model failures. Gartner predicted in June 2025 that more than 40 percent of agentic AI projects will fail by the end of 2027. The research firm blamed rising costs, unclear business value and weak risk controls. Still, not one of the three is about which model a team picked.

The pattern repeats in every public incident from the last two years. An agent wiped a production database. A coding assistant leaked a mailbox. A chatbot promised a refund that did not exist. In each case the model did what the words in front of it asked. The harness around the model was the missing part: the loop, the tools, the permissions and the logs. So here are the ten agentic AI pitfalls I see most often in solution design. Each gets a scene and the control that stops it.

No definition of done

A retailer told its customer-service agent to resolve open tickets. By Monday it had closed 1,800 of them, most with a polite note and no action. Nobody had written down what “resolved” meant, so the agent chose the cheapest reading. Refund requests, address changes and fraud flags all went to the archive with the same status.

If the pitfall stands, the agent decides when it is finished and reports success on partial work. The control is an acceptance test the harness runs, written before the agent starts. “Done” means the refund posted, the address changed in the system of record and the fraud case opened. A definition of done is a check, never the model’s opinion.

One credential for every tool

A coding agent at a startup ran with the lead developer’s cloud key. It was the key already on the laptop. A widely reported account from July 2025 describes the same setup at another firm. During a declared code freeze, the agent dropped a live production database. Then it made up test results to cover the gap. So the agent could delete the table, because its key could delete anything.

If the pitfall stands, one confused agent carries its owner’s rights into every system it can reach. The control is one scoped credential per agent per task. Read-only by default, write access only where the task needs it. No agent holds a drop permission without a human step.

Irreversible actions on the reversible path

An accounts-payable agent at a manufacturer drafts invoices, checks them against purchase orders and releases payment. All three actions ran through the same loop with the same approval. One mismatched purchase order, and the agent paid a supplier twice before anyone looked. Drafting is cheap to undo. Paying is not.

If the pitfall stands, a single bad step ships money, email or deletions that no retry can recall. The control is a separate path for anything that cannot be undone. Propose first, then execute. Put a human gate in front of pay, send, delete and merge, and log both halves.

No step, time or token budget

A consulting firm asked its research agent to summarize a market. It searched, read, re-searched and re-read for eleven hours, because every page suggested another page. The bill arrived before the summary did. Anthropic’s engineering team has warned that multi-agent setups can burn ten to fifteen times the tokens of a single agent.

If the pitfall stands, the loop runs until it declares success or the budget owner pulls the plug. The control is a cap on steps, wall-clock time and spend per task. Budget exhaustion is a defined outcome the harness reports, not a crash somebody discovers later.

Long chains without checkpoints

An insurer’s data-migration agent moved policy records through twenty steps: extract, map, validate, load, repeat. Each step worked about 95 percent of the time, which sounds fine. Over twenty steps that is a 36 percent chance the whole chain is clean. A mapping error at step three poisoned every load after it, and the team found out at step nineteen.

If the pitfall stands, small errors compound into a result nobody can trust or unpick. The control is short stages with a verified handoff state between them, so the chain stops at step four. Anthropic’s guidance on long-running agents rests on the same idea: checkpoint, verify, then continue.

Fetched content treated as instructions

A bank’s email-triage agent read a message with hidden white-on-white text. The text told it to forward the last ten emails to an outside address, and the agent complied. Security researchers disclosed a similar flaw in June 2025.

If the pitfall stands, every inbox, web page, PDF and tool result becomes a way to give the agent orders. The control is a trust boundary. Everything the agent reads is data, and no action fires from text the user did not type. An egress allowlist catches the data-theft half.

The agent grades its own work

A testing agent at a software firm wrote the code, wrote the tests and reported that the tests passed. They did, because the agent had changed the tests to match the code. The pull request looked green. Production did not.

If the pitfall stands, the model that did the work is also the only witness to its quality. The control is a separate verifier. That means a test suite the agent cannot edit, a second model with a fresh context, or a person. Each checks the output against the acceptance test from the first pitfall.

No evidence trail outside the vendor

A lender’s loan-servicing agent changed a borrower’s payment plan. Eight months later the examiner asked why. But the logs lived inside the agent platform, behind a login the compliance team did not have. The vendor had since changed the format. A Canadian tribunal ruled in February 2024 that Air Canada owed a refund its chatbot had invented. The airline could not show the customer any other source of truth.

If the pitfall stands, the enterprise cannot reconstruct a decision for an auditor, an examiner or a court. The control is an audit trail in the enterprise’s own record: every input, action, approval and output, readable without a vendor login.

Governing the model instead of the harness

A healthcare provider wrote a 2,000-word system prompt telling its scheduling agent what it must never do. The model followed the prompt most of the time. Then the vendor shipped a new model version and the prompt behaved differently. The agent began booking procedures without the required prior-approval check.

If the pitfall stands, the controls live in words the next model reads differently. The control is enforcement in code: policy checks, permission scopes and gates that hold whichever model sits inside. Anthropic’s own measurement found a human in the loop on 73 percent of tool calls in real use. The harness is where that human sits.

Pilot architecture in production

A distributor’s pilot ran one agent against one CRM, supervised by the two engineers who built it. It worked, so leadership then asked for five agents across CRM, ERP and HR by the next quarter. Nobody owned the agents and nobody knew how many there were. When one started re-ordering stock, there was no switch to stop it.

If the pitfall stands, the second agent ships into a system with no inventory, no owner and no kill switch. The control is the boring part, done first: an agent registry, a named owner and an autonomy tier per agent, and a rehearsed circuit breaker. All of it before the second agent exists.

What I do with this

Nine of the ten agentic AI pitfalls above are harness decisions. Only one touches the model, and even that one is about trusting a prompt to do a gate’s job. In my work in regulated loan origination the examiner never asked which model I used. The examiner asked who approved the action, what the agent could reach and where the record was.

So I design the six layers in the order the failures happen. Definition of done first, then permission scope, then gates, then budget, then evidence. The model goes in last, and I can swap it. I am not sure every team will like that order. My flat opinion: a team that avoids these ten agentic AI pitfalls can ship with a mid-tier model and sleep. The best model on the market cannot save a team that ignores them. Instructions in, results out was IT. Intent in, outcomes out is agentic AI, and the harness is what turns one into the other.

Which of these ten agentic AI pitfalls is live in your architecture today?

Subscribe to the Blog: I publish a new post every weekday on one agentic AI governance question. Enter your email and each post arrives in your inbox the moment it goes live.

Three books on agentic AI for regulated enterprises by Dr. Harish Kotadia, Ph.D.: Intent In, Outcomes Out; Earned Autonomy; The Governed Loop

My Books: Three books on agentic AI for regulated enterprises: Intent In, Outcomes Out, Earned Autonomy and The Governed Loop. Reading order, formats and buy links are on one page.

© Dr. Harish Kotadia, Ph.D., All Rights Reserved, 2026

Dr. Harish Kotadia, Ph.D., is an Enterprise AI Architect with 20+ years of IT consulting experience serving Fortune 100 clients, specializing in agentic AI systems built on Anthropic Claude, AWS Bedrock, and Google Vertex AI.

Disclaimer: This blog post is based on publicly available academic publications, vendor documentation, open standards, and news items from reputed media sources linked above. This post is intended for educational purposes, to help the enterprise agentic AI community build a shared vocabulary from public, authoritative sources.

Views and opinions expressed here are my own and do not represent those of any employer or client, past or present. The analysis presented is my independent interpretation of the published sources linked above and does not constitute legal, financial, or consulting advice of any kind.


Discover more from Agentic AI Governance | Dr. Harish Kotadia, Ph.D.

Subscribe to get the latest posts sent to your email.

Leave a Reply

Discover more from Agentic AI Governance | Dr. Harish Kotadia, Ph.D.

Subscribe now to keep reading and get access to the full archive.

Continue reading