The Testing Academy · Class Notes Saturday, 29 August (IST)
Live class · study guide

BLAST, and the first agent that runs without you

AI Tester Blueprint 4xSaturday, 29 August (IST)2h live

Level three. Prompts were level one and skills were level two; an agent is the thing that keeps running when you close the laptop. This session gave the agent a build discipline, BLAST, which forces the AI to write down its plan, findings and progress as files you can read, and then built two working Jira agents in n8n.

By Pramod Dutta, The Testing Academy. Study notes from the live AI Tester Blueprint 4x class, rebuilt from the session recording. The demo API keys shown on screen during the class were deleted live and are deliberately not reproduced here. Pricing and model names are recorded as stated in the session and move quickly, so treat them as the shape of the answer rather than a quote.

01

Where this sits

Three levels, and the batch has now done all three.

Level What it is Runs when
Prompt One instruction, typed You type it
Skill A reusable, structured template You invoke it
Agent A brain with memory and tools On its own, 24/7

The line drawn in class about the first two: prompts and skills are dead until you open them. They live inside your Copilot or Claude Code window and do nothing while you sleep. An agent is deployed, reachable, and running.

Large language model brain cannot act, cannot react AI agent brainmemorytools acts on its own, remembers, reaches other systems The shorthand from class: an LLM is B. An agent is B plus M plus T.
Adding memory and tools is what turns an answer into a worker.
02

BLAST, and the problem it solves

The batch had already built a test plan creator from a single prompt. It worked. The question that opened this session was harder: what did it actually do? What did it find, what did it decide, what did it try first and abandon? Nothing was written down, so nobody could say.

That gap is not academic. The hiring point made in class: candidates are being rejected for using Copilot without being able to explain the architecture it produced. Do not be a blind AI user.

BLAST is a build discipline that forces the AI to show its working.

Letter Phase
B Blueprint, plan it first
L Link, find the sources and references it needs
A Architect, structure it properly
S Style, tidy and finish
T Trigger, deploy and run it
03

Phase 0, the part people skip

Before any of the five letters, BLAST mandates a set of markdown files, and this is the whole trick:

File Holds
task.md What we are building
findings.md What the AI discovered along the way
progress.md What it did, timestamped
llm.md Which model, and any model-specific notes
prompt.md The prompts actually used
One prompt, no framework a prompt an app and no record of how, so you cannot explain it Same prompt, with BLAST a prompt an app task.mdfindings.mdprogress.mdprompt.mdllm.md plus a record you can read and defend
Same output, plus the paper trail. The paper trail is the point.

progress.md is timestamped, entry by entry. When Groq hit its rate limit mid-demo and the build switched to DeepSeek, the progress file recorded the switch and the reason. That is the whole argument for BLAST in one accident: something went wrong live, and the log explained itself without anyone reconstructing it afterwards.

BLAST is a template, not a tool. It forces whichever assistant you use, Copilot, Claude Code, Codex, to follow phase 0 through phase 4 rather than jumping straight to code. The same idea as RICEPORT for prompting, applied to building.

04

n8n, and the three letters that matter

The second half moved to n8n, a visual workflow builder. The framing was deliberate: the tool does not matter, the concept does. The same ideas work in Langflow or anything similar.

The concept is TAR:

TAR TriggerActionResult what starts it: a message,the agent works: reads thewhat comes back: a ticket, a schedule, a webhookbrain, calls its toolsa summary, a Slack post Every agent you will build is these three boxes. Learn the shape, not the buttons.
Trigger, action, result. The tool changes; this does not.

Two agents were built live, both against Jira: one that fetches a ticket, one that creates a bug. Each needed three pieces wired together: a model credential, a Jira credential, and simple memory.

The part worth noticing: once published, the agent has a public URL and is reachable around the clock. That is the difference from a Copilot workflow, which only exists while you have the window open.

05

Choosing a brain, and what it costs

The session was explicit that the model is interchangeable.

Option Situation
Groq Free, and rate limited. It hit the limit live during this class.
DeepSeek Around five cents per million tokens, as quoted in class
OpenRouter Needs roughly a five dollar minimum
Ollama, LM Studio Local, free, your own hardware
Claude, GPT Paid API, billed separately
n8n cloud trial Fifteen days, with GPT-5 mini credits included, which is what the class used

Two cost points worth keeping:

  • Generating the plan is a one-time spend. Once the agent is deployed, the recurring cost is the small per-run token use, not the build.
  • Use small models for the repetitive parts. The figure given for a real suite: around a hundred test cases running daily, in the region of ten to twenty dollars a day, using mini models for the routine work.

The rule that matters more than any of the above. Do not use Groq, or any personal API key, on a company laptop without permission. Ask first, then use it. The safest key is always the one your employer gave you. Demo keys shown on screen in class were deleted immediately afterwards, and none are reproduced in these notes.

06

What comes next

The test plan produced today came from one input: the ticket. The obvious improvement is to give the agent more to read: your existing regression suite, your Selenium or Playwright repo, your team's documents and PDFs.

That is RAG, and it is next week's session. It is also the answer to the coverage question raised in class, where a plan built from a ticket alone reaches only about thirty per cent of what the team actually knows.

07

Tasks and announcements

  • Homework: build a test plan, test case or test strategy agent in n8n. One agent, yours, working end to end.
  • The 60-day branding challenge continues. Two to three LinkedIn posts a week, plus blogs, Medium and Instagram if you can. The reason given, plainly: everyone now has Copilot, so the differentiator is what you can show and explain.
  • The BLAST prompt will be shared as prompt.md in the class repository.
  • Tomorrow: the n8n session continues with the test plan generation walkthrough.
  • Next week: RAG, and the QA Copilot project.
  • Related: prompt engineering for QA, and your first CrewAI agent from the 3x batch.