Story 1: Orchestra Without a Conductor
On purpose, architecture, and the question that stays on the whiteboard
The room has no windows.
Not because it’s hidden. Time there simply doesn’t exist the way it does outside. It’s like a server room running processes nobody sees until they break.
Four people are sitting in this room.
* * *
Harrison and the Map
Harrison Chase, founder of LangChain, has a whiteboard in front of him. It’s covered in arrows. Many arrows. Some pointing left, some right, some looping back.
“This is orchestration,” he says, gesturing at the tangle of lines.
Someone from the corner replies: “That looks like the London Underground.”
“Exactly,” Harrison says, and smiles. “And it works the same way. Every line has its own logic. Every agent knows its stop. The problem starts when someone doesn’t know where they want to go.”
Silence.
“The map doesn’t determine the destination,” he adds, quieter. “The map just makes sure you don’t miss your stop.”
That’s the first tension moving through the room: LangChain is an architecture of movement, not an architecture of purpose.
* * *
Andrew and the Question Nobody Asks
Andrew Ng doesn’t sit. He paces, the way he always does when he lectures — because he’s always lecturing, even when he’s just talking.
“Understand,” he says, “that the biggest mistake companies make is deploying an agent and waiting for a miracle. An agent isn’t a miracle. An agent is iteration.”
He takes a marker and writes on the board: Draft → Reflect → Revise → Reflect → …
“This is an agentic workflow. Not a linear chain - a loop. The agent writes, evaluates itself, revises, evaluates again.”
“But evaluates against what?”
Andrew stops.
“Against the instructions you give it,” he answers.
“And if the instructions are wrong?”
A long pause.
“Then the agent will do the wrong thing exceptionally well.”
* * *
Kambhampati Enters Through the Side Door
Subbarao Kambhampati doesn’t wait for an invitation. He enters as if he had a prior arrangement with the truth.
A professor at Arizona State University, one of the sharpest critics of what he calls the “planning illusion” in language models.
He sits across from Andrew.
“Iteration isn’t thinking,” he begins, without preamble. “When an LLM generates the next token, it doesn’t plan, it samples. And sampling that looks like planning isn’t the same as planning.”
“But the results - ”, Andrew starts.
“The results are good as long as someone else already solved the problem in the training data. When the task becomes genuinely new, the LLM improvises. And improvisation without understanding is sophisticated guessing.”
Harrison looks at his map. Andrew holds the marker.
“So orchestration doesn’t work?” someone asks.
“Orchestration works,” Kambhampati says. “The question is what it’s orchestrating. Agents that don’t understand, or agents that have enough context that they don’t need to?”
* * *
Lakhani and the Strategic Contract
Karim Lakhani arrives last. Harvard Business School. Author of a book on how AI reshapes organizations from the inside.
He doesn’t sit right away. He stands by the window - which isn’t there - and stares at the space where a window should be.
“You’re talking about how agents work,” he says. “I’m asking why they exist at all.”
“Because they increase productivity,” Andrew answers.
“Productivity of what? For whom? In service of which value?”
Silence.
“Every organization that deploys agents implicitly signs a contract with the future. But most organizations don’t read the contract. They just sign it.”
Lakhani takes the marker from Andrew’s hands. He writes two words on the board: Exploitation and Exploration.
“Agents are excellent at Exploitation: optimizing what we already know. They’re poor at Exploration: finding what we don’t yet know. And most organizations want Exploration, but pay for Exploitation.”
* * *
Soft Orchestration
All four are sitting quietly now.
And then someone asks the question nobody had spoken aloud:
“What if orchestration isn’t in the code?”
Workflow orchestration is a score: every note determined in advance. Semantic orchestration is a jazz standard: the structure exists, but within it there’s room for judgment.
But soft orchestration, the kind this story is searching for, is the culture of an ensemble. It isn’t in the notes. It isn’t in the standard. It’s in that moment when the musician knows, without being told, that now is the time for silence.
You can’t program that. But you can create the conditions in which it emerges.
* * *
The One Missing from the Room
There’s no business architect in this room.
A business architect isn’t the one who knows how agents work. They’re the one who knows why the company exists - and whether the orchestration supports that why, or quietly erodes it.
They’re not in the room because nobody invited them.
* * *
Agents orchestrate processes.
Processes are orchestrated by rules.
Rules are orchestrated by values.
Values are orchestrated by purpose.
* * *
And that - more than any hallucination, more than any misaligned prompt - is the real risk of agentic AI.
Not that the orchestra plays wrong notes.
That the orchestra plays the wrong piece.
* * *
Purpose cannot be orchestrated by anyone but you.
* * *
Author’s Note
The characters in these stories are real thinkers whose published work inspired these dialogues. The conversations are fiction. The ideas build on their foundations — the synthesis is mine.
Harrison Chase, Andrew Ng, Subbarao Kambhampati, and Karim Lakhani are publicly accessible through essays, lectures, and books. Ana Kovac is a fictional customer — and every customer your system knows only through data.