---
title: "Dr Srinivas Padmanabhuni (@srinivaspadmanabhuni) | Hashnode"
description: "AI assurance specialist. PhD in AI (University of Alberta), past President of ACM India, TEDx speaker. Writes the \"Testing Agentic AI\" series and publishes t..."
url: "https://hashnode.com/@srinivaspadmanabhuni"
---

@srinivaspadmanabhuni

# Dr Srinivas Padmanabhuni

@srinivaspadmanabhuni · Joined October 2026

[Website](https://sites.google.com/view/srinivaspadmanabhuni)

## Comments

[The refund example is the strongest part: issuing a refund without approval is a release blocker regardless of the overall score. That is also where traces matter. In one run I reviewed, an agent sent a DELETE through a read-only tool. The server refused it, but nothing in the recorded trace marked the step as denied, so a checker reading the trace had no signal to flag. Policy checks before agent actions help a release review when the refusal is written into the trace as well. I check recorded agent traces with AiEnsured agentest, 165 tests: pip install agentest-ai Developers are welcome to enhance and contribute to it.](https://hashnode.com/posts/how-to-evaluate-ai-systems-before-production/6affecab9b82e9f4dedba4ba/comment/6b017f53fb0672e6a7ab1be7)

Comment · Article · 1h ago · [How to Evaluate AI Systems Before Production](https://hashnode.com/posts/how-to-evaluate-ai-systems-before-production/6affecab9b82e9f4dedba4ba)

[The line that consensus is not the same as correctness is the one I'd stress too, because sixteen agents agreeing can look more convincing than one agent guessing. That makes the trace the thing worth testing, not only the final answer. In AiEnsured agentest, an Agentic AI Testing platform with 165 tests, one failure from its report is a shopping agent that told the user an item "has been added to the cart" while the cart was empty, after it requested checkout at a price that came from nowhere. A cross-check between agents would not catch that unless something verifies the actual tool state. Developers can try it with the command below, in text only. pip install agentest-ai](https://hashnode.com/posts/perplexity-put-16-agents-on-one-task-the-cross-check-is-the-point/6b00fddd108976d53e819da7/comment/6b0157f269c11d9c3bec16d7)

Comment · Article · 4h ago · [Perplexity Put 16 Agents on One Task. The Cross-Check Is the Point.](https://hashnode.com/posts/perplexity-put-16-agents-on-one-task-the-cross-check-is-the-point/6b00fddd108976d53e819da7)

[The line that stood out to me is the "start with no done" ledger state: a crashed run looks like a run that never started. The same gap shows up in agents, where the final message says something succeeded and nothing independent confirms it. A travel-planning agent we tested replanned from Whitefield to Koramangala, found an inn at INR 4800, then tried hold\_booking at INR 5900, a price no search had returned. The tool refused, but the flight hold still went through and the final answer said "Infeasible". Only the recorded trace showed a half-completed side effect. So the ledger needs to come from what actually happened, not from the agent's summary. That's what we built AiEnsured agentest, an Agentic AI Testing platform, to judge from recorded traces. It has 165 tests: pip install agentest-ai Developers are welcome to enhance and contribute to it.](https://hashnode.com/posts/safe-to-run-twice-idempotency-for-small-assistant-jobs/6b014420832dafbfcf00a61a/comment/6b014fff69c11d9c3bec0097)

Comment · Article · 4h ago · [Safe to run twice: idempotency for small assistant jobs](https://hashnode.com/posts/safe-to-run-twice-idempotency-for-small-assistant-jobs/6b014420832dafbfcf00a61a)

[Good point that schema validation checks structure but not whether the product exists or the user is allowed to see it; that gap is easy to miss. A check I'd add to the workflow test is reading the saved trace afterwards: does the assistant's final answer match what the tool really returned, for example saying a quotation was prepared when prepare\_quotation errored? AiEnsured agentest runs that kind of trace check (pip install agentest-ai, 165 tests), and developers are welcome to enhance and contribute.](https://hashnode.com/posts/how-to-build-reliable-ai-function-calling-for-business-applications/6afefb8af9d7c9d4ae00c8a2/comment/6aff094229c4bc505b3f915e)

Comment · Article · 1d ago · [How to Build Reliable AI Function Calling for Business Applications](https://hashnode.com/posts/how-to-build-reliable-ai-function-calling-for-business-applications/6afefb8af9d7c9d4ae00c8a2)

[A readiness gate that only asks whether each tool call was allowed can still pass a bad run. I looked at a travel-planner agent where the guard refused a hotel hold because the price didn't match any search result, yet the flight hold had already gone through, and the final answer said "Infeasible". Half a trip was held. Checking the final state after a failed step is the gate I'd add.](https://hashnode.com/posts/testing-ai-agents-before-deployment/6afd7281c3b20584bb31a192/comment/6afed44d6cda2eb6b3706fe5)

Comment · Article · 2d ago · [Testing AI Agents Before Deployment](https://hashnode.com/posts/testing-ai-agents-before-deployment/6afd7281c3b20584bb31a192)

[The refund example where schema-valid JSON cites a knowledge-base chunk that retrieval never returned is a sharp way to show that shape and truth are different checks. Your rule that the same error twice means stop resonates: in one shopping-agent trace I looked at, the agent called open\_url with the identical search URL more than ten times and ended with an empty answer. Checking the tool-call sequence, not just the final output, is what exposed that retry storm.](https://hashnode.com/posts/structured-llm-outputs-pydantic-django/6afcd459c7d39edf63adb6b0/comment/6afd43459a49bb4f2637db4c)

Comment · Article · 3d ago · 1 · [Valid JSON, Wrong Refund: Structured LLM Outputs in Django (Part 7)](https://hashnode.com/posts/structured-llm-outputs-pydantic-django/6afcd459c7d39edf63adb6b0)

[I like the separation between deterministic checks, source-backed fact checks, and rubric-based model judgment; the warning that test mode must redirect write actions, not just prompt the agent, is especially important. In one travel-planner run I examined, search returned Koramangala Inn at INR 4,800, but the agent requested a hold at INR 5,900; the guard refused that hotel hold only after the flight hold had already succeeded, leaving the trip partially held. That is why I think action-level and workflow-level checks matter: AiEnsured agentest, an Agentic AI Testing platform, is built to surface failures even when a guard catches one step after an earlier action has already happened.](https://hashnode.com/posts/ai-agent-evaluation-testing-2026/6afc5e0f1b79ed44faeab615/comment/6afc710a48415021cf24a4e9)

Comment · Article · 3d ago · [Testing AI Agents: How to Measure the Quality of Your Automation](https://hashnode.com/posts/ai-agent-evaluation-testing-2026/6afc5e0f1b79ed44faeab615)

[The replay method is strongest when it separates model drift from retrieval drift: reusing the same retrieved context makes the comparison fair, but can miss a changed retrieval path. In my work on AiEnsured agentest, an Agentic AI Testing platform, an a17 multi-hop RAG run retrieved the documents containing the answer yet still replied, “Not found in the official documents.” That’s why I’d pair outcome agreement with citation and retrieval checks during disagreement review; your stratified sampling and reviewer-corrected cases offer a practical place to do it.](https://hashnode.com/posts/how-to-upgrade-your-llm-without-breaking-production-a-replay-testing-guide-with-zizkadb/6afc17497501f8e04b309594/comment/6afc5971c665382e7b18beec)

Comment · Article · 3d ago · 1 · [How to Upgrade Your LLM Without Breaking Production: A Replay Testing Guide with ZizkaDB](https://hashnode.com/posts/how-to-upgrade-your-llm-without-breaking-production-a-replay-testing-guide-with-zizkadb/6afc17497501f8e04b309594)

[Useful that the rollback also does git checkout, because pruning the chat history alone would leave the failed patch sitting in the workspace. Two checks I'd add to the 20-case suite: run the same task where the first failure is a misleading error (a test that fails for an environment reason, not because of the patch) and see whether the 927-character lesson teaches the agent a wrong rule that then hurts the retry. Also compare against a task needing three or more attempts, where lessons accumulate, to see if the lesson log itself starts poisoning the prompt. A single run per mode on gemini flash is a thin sample, so I'd repeat each mode ten times and report the spread.](https://hashnode.com/posts/e19-error-trace-pruning/6ae6d2737a5b1b29b641f340/comment/6ae7129eec361e7786cefbee)

Comment · Article · Nov 1 · [Why Keeping Failed Code Attempts in Chat History Poisons Your Coding Agent: Error-Trace Pruning vs. Accumulated History](https://hashnode.com/posts/e19-error-trace-pruning/6ae6d2737a5b1b29b641f340)

[The G2 case, a 200 with an error banner, shows why I'd make the agent's 'done' a read-back rather than a UI phrase. After each submit, have it re-fetch the product page or dashboard row and compare the fields it entered. That would also have caught the blank Vendor Name after the domain lookup: the lookup overwrote a field the agent had already filled. For the refused declarations like 'I am 16 or older', I'd keep a small regression set of forms that deliberately contain such checkboxes and assert the run halts there with the field named in the log. Behaviour that only appears when a form asks something unusual is easy to lose in a prompt tweak.](https://hashnode.com/posts/what-broke-when-an-ai-agent-filled-in-50-sign-up-forms-for-us/6ae542f999f8168a318e9e03/comment/6ae5c53709eb1df3dfb388cf)

Comment · Article · Oct 31 · [What broke when an AI agent filled in 50 sign-up forms for us](https://hashnode.com/posts/what-broke-when-an-ai-agent-filled-in-50-sign-up-forms-for-us/6ae542f999f8168a318e9e03)
