Slide 1 of 29: Title
← MSRITFACULTY DEVELOPMENT PROGRAMME · RAMAIAH INSTITUTE OF TECHNOLOGY, BENGALURU
AGENTIC AI IN PRACTICE: BUILDING AND ORCHESTRATING INTELLIGENT AGENTS, DAY 1 · 13 JULY 2026
Agentic AI in theReal World.
Orchestrating agents and humans at work.
SHASHANK MANJUNATH
PRODUCT ENGINEERING LEADER · WRITES AT BYSHASHANK.COM
I · THE OPENING STORY
11 days
One simple screen. Two days of building. Nine days of questions, crossing oceans, one at a time.
THE ARITHMETIC OF ELEVEN DAYS
THE BUILDING TOOK TWO DAYS. THE LOOP TOOK THE OTHER NINE.
THE SAME HANDOFF, TODAY
One conversation.
An agent reads the document before any developer sees it, finds the questions, and gets them answered, on day one.
HOLD THAT GAP. IT'S WHAT THIS WEEK IS ABOUT.
WHAT THIS HOUR IS, AND ISN'T
No frameworks.
No code.
No hype.
You have four days of genuine experts for that. This is one practitioner's honest account: what worked, what surprised us, and what it means for the people who will teach the next generation of engineers.
II · HOW WE GOT TO "AGENTIC"
WHAT AN AGENT ACTUALLY DOES
Plans
Sequences steps toward a goal, instead of waiting for the next instruction.
Uses Tools
Searches a codebase, calls an API, reads a design file, writes a document.
Remembers
Keeps track of where it is in the work, across steps.
Decides
Chooses what to do next, without a human queuing up the next command.
A BRILLIANT CONVERSATIONALIST, UNTIL IT WASN'T BOLTED TO A CHAIR.
THE HONEST DEFINITION
Autonomy is a dial,not a switch.
The question is never ‘can we build an agent that does X?’ We nearly always can, now. The question is: how far do we turn the dial (for this task, with these stakes) and what do we build around it?
CARRY THIS QUESTION INTO THURSDAY (DEPLOYMENT) AND FRIDAY (ETHICS).
WHY NOW: A THRESHOLD CROSSING
Long Context
Models that can hold enough of your world in their head to reason over it.
Tool Use
The model can act: search, fetch, execute, and read the result. Language became action.
Cheap Orchestration
One engineer, one laptop, one afternoon. You'll do it yourselves this week.
None of these is magic. But thresholds matter. All three crossed at once.
III · THE HUMAN PROBLEM
A dozen people.Six time zones.
Every PM wrote differently: three tight paragraphs, or nine pages, or bullet points that made sense only to her. The quality of what reached a developer depended on who wrote it, and how tired they were that week.
Every handoff leaks.
A constraint the designer fought for ("no delete button on this screen") quietly vanished between three hands and two documents.
A decision made in chat was buried in a week, and re-made six weeks later, the other way.
The quality of a spec depended on who wrote it, and how tired they were.
A DOZEN PEOPLE · SIX TIME ZONES · EVERY BOUNDARY A LEAK
THE REFRAME
Nobody was failing.The handoffs were.
Everyone did locally sensible work, and information decayed at every boundary between them. The team didn't lack intelligence. It lacked process memory.
YOU CAN'T FIX 'BE MORE CAREFUL.' YOU CAN FIX 'INFORMATION DIES HERE.'
THE ONE SENTENCE TO STEAL
We didn’t ask "where can we add AI?"
We asked: "where does information die?"
Walk your own department's chains: between you and your scholars, between one committee and the next. That's where this technology belongs in your world.
IV · WHAT WE ACTUALLY BUILT
"Run the pipeline."
Four words from a product manager. One orchestrating agent runs the relay race that used to take days, and brings back only the genuinely human questions, answered like a colleague on Slack, written into the documents, not the chat.
THE RELAY, STEP BY STEP
ONE CONVERSATION. THE ELEVEN DAYS HAPPEN IN AN AFTERNOON.
WHAT STAYS HUMAN
The PM still writesthe product thinking.
What should this screen do, who can see it, what must it never do. The agent expands it, audits it, questions it, but the thinking, and the answers, land in the documents. Never the chat.
THE THESIS
This week you'll learn to build agents in code.In my team, the orchestration language is English.
The orchestrator is a document. The gates are documents. The checklists are documents: plain English, in version control. Improving the system means editing prose, like amending a constitution.
WHAT THAT MEANS
The orchestrator itself is a document: a long, careful piece of English prose that tells an agent what to read, when to stop, what it must never do.
The quality gates are documents. The checklists are documents. The templates are documents.
Improving the system means editing prose and committing it, like amending a constitution, not deploying a service.
The scarce skill is precise writing: constraints and judgment, clear enough for a machine to follow and a colleague to audit.
NOT PRIMARILY PROGRAMMING. PRECISE WRITING.
THREE DECISIONS · THREE SCARS
Fresh Eyes, Never The Author
Every spec our generator writes is audited by a second agent with clean context (no memory of the writing) and for high-stakes screens, a genuinely different model. An author reviewing its own work has the author's blind spots. You don't referee your own paper.
THE WAR STORY
A spec that was internally flawless: consistent, complete, beautiful. Because the reviewer checks the actual codebase, not just the document, it found a half-built version of the same screen already sitting in the repo, structured completely differently.
Review must reach reality, not just the artifact.
THREE DECISIONS · THREE SCARS
Hard Stops, Not Helpful Guesses
Early on, we let the agent be helpful when things went wrong: a design link that didn't resolve, and it would improvise, guess the screen's name, scaffold something plausible, carry on cheerfully.
THE WAR STORY
It felt efficient. It was a disaster in slow motion: everything downstream inherited the guess, dressed up in confident, professional-sounding prose.
Some failures should be loud. If the input can't be verified, stop the entire process. Create nothing. Name what failed, and wait for a human.
“The most dangerous agent is not an incapable one. It's a helpful one with wrong information.”
IT FAILS FLUENTLY, AND THE ERROR COMPOUNDS DOWNSTREAM, WEARING A SUIT.
THREE DECISIONS · THREE SCARS
Humans Author Decisions
Our agent has standing orders: it never invents a product decision. When it hits a genuine judgment call (should this action be irreversible? who is allowed to see this data?) it marks the question and routes it to whoever actually owns that call.
THE WAR STORY
If the PM ventures a guess about someone else's call, the agent records it as provisional, to be confirmed by the owner, never laundering a guess into a decision.
Autonomy for logistics; humans for judgment. We set that dial deliberately, task by task, not globally.
HOW WE KNOW IT WORKS
Clarification round-trips per feature.
Not accuracy. Not tokens. Not latency. How many times a developer had to stop and ask a question the spec should have answered: counted on fingers, no tooling required.
Measure outcomes, not artifacts.
THE PRACTICE
Once a month.
I sit with the log and ask one question: which single prompt or template change would have removed the most round-trips? We make that one change. Just one. Then we measure again.
NO GRAND REDESIGNS. ONE EVIDENCE-DRIVEN EDIT A MONTH, COMPOUNDING.
V · LESSONS THAT TRAVEL
FOR THIS ROOM, SPECIFICALLY
Machines generate.Humans constrain.
BEFORE GENERATION
Specification: saying precisely what you want, and what must NOT be done. The most load-bearing artifact in our pipeline is the human-written list of must-nots.
AFTER GENERATION
Judgment: reviewing critically with reality in hand, knowing what to check mechanically and what to check with your own head.
The middle (generation) is what our curricula spend most of their hours teaching. It's also what the machines are eating fastest.
VI · YOUR WEEK: I'VE GIVEN YOU A MAP; THEY'LL GIVE YOU THE TERRITORY
Bring them your hardest questions, and bring your own institution's workflows into every session.
THE SEND-OFF
The people didn't get smarter.The handoffs did.
Same people. Same time zones. Same oceans.
THANK YOU
Build agents this week. But remember what you're really building:
a new way for people to work together.
QUESTIONS WELCOME
SHASHANK MANJUNATH · BYSHASHANK.COM
byshashank.