PSG · 11 August 2026 · 60 min · 6 live demos

Building Agentic AI
Systems & Multi-Agent Architecture

Built live in Colab · why the hard part was never the model

Siddarth Kengadaran Experience Strategist

Foundations

A Model Call Is Not an Agent

A call is a function. Text in, text out. It cannot look anything up, and it cannot take a second step.

TOOLS reaches past its own context FROM YOU TASK THINK ACT OBSERVE GOAL MET? YES ANSWER NOT YET A LOOP — IT GOES AGAIN

Tools

It can reach past its own context and act on something real.

A Loop

It can take a next step based on what it just learned.

A Goal

It decides when it is finished. You did not write the stopping condition.

Those three, and nothing else. Every framework you will read about this year is packaging around them.

Foundations

One Question Separates Them

Who picks the next step?

ONE TURN ASK MODEL ANSWER

Plain call

Nobody. There is no next step. One turn, then it stops.

YOU WROTE THIS ORDER STEP 1 STEP 2 STEP 3

Workflow

You do, in code. You wrote the sequence and you know it in advance.

NOT KNOWN IN ADVANCE STEP STEP STEP

Agent

The model does. You do not know the sequence, because it depends on what it finds.

This is the only line that matters. Everything else — frameworks, diagrams, vendor names — sits on top of it.

Live · Colab

Watch it choose its own next step

Notebook §1  ·  same question, once without tools and once with two

Without tools: “I don’t have access to your analytics.” Correct, and useless.
With tools: it lists what exists, fetches what it needs, then answers.
Nobody wrote that sequence. That is the whole definition, on screen.

Foundations

Most Things Should Not Be an Agent

If you can write the steps down, write them down. A for loop is cheaper, faster, debuggable, and it will not invent a fourth step at two in the morning.

Reach for an agent when the sequence depends on what you find. Our second tool call was decided by the answer to the first. Rename those metrics and the plan changes, with no code edit.

That flexibility is the entire purchase. The rest of this hour is what it costs.

Foundations

Three Honest Reasons for a Second Agent

One agent with tools just solved that on its own. So why add another?

Separate context

One agent reads fifty documents and hands forward a page. The next never carries the fifty.

Separate permissions

The agent that reads the database should not be the agent that sends email. That is a security boundary.

Parallelism

Ten independent files, ten agents, one wall clock.

Notice what is not on that list: “so each one can specialise in a role.” That is the reason almost everyone actually reaches for, and it is the weakest one.

Opening

The First Thing Everyone Builds

You have a hard task. You give it to one agent. It does a mediocre job. So you do the obvious thing.

01

Add a researcher agent to gather the facts.

02

Add a writer agent to draft from what the researcher found.

03

Add a critic agent to review the draft.

04

Add a manager agent, because now somebody has to coordinate the other three.

BRIEF DRAFT REVIEW TASK RESEARCHER WRITER CRITIC MANAGER ANSWER EVERY ARROW IS PROSE, AND EVERY ARROW REPACKAGES THE CONTEXT

Four agents. It is slower, costs roughly four times as much, and the answer is often worse.

Why That Happens

You Rebuilt a Meeting

Every arrow between two agents is a place where context gets repackaged, summarised, and paraphrased. The task survives the trip. The understanding of the task does not.

You have met this before. Brooks wrote it down in 1975: adding people to a late project makes it later, because the communication paths grow faster than the workers do.

Agents are not exempt. They are worse, because they never say "wait, I don't follow."

The Tell

Ask a four-agent system why it produced a particular answer.

You get a plausible reconstruction, not a trace. Nobody kept the evidence, because the handoffs were prose, and prose does not carry proof.

An answer nobody can account for is not an answer. It is output.

1 AGENT · 0 PATHS 2 · 1 PATH 3 · 3 PATHS 4 · 6 PATHS

Part I · 15 min · demos A, B

The Real Bottleneck

Live · Colab

Build it, then measure it

Notebook §3 – §4  ·  one agent, then four in a chain

The answer is fine. All four facts, sensible advice, nothing lost.
So was the single agent’s. You could not pick them apart.
Four agents cost ~5× the tokens and ~5× the wait — for that.

The Reframe

The bottleneck is not model capability and it is not agent count. It is contextual transaction cost — what it costs to make task context usable across a boundary.

THE BOUNDARY AGENT A AGENT B … OR A HUMAN, OR THE NEXT RUN HANDOFF COMPRESSION DRIFT VERIFICATION TOKENS GOVERNANCE

Every boundary charges a toll. Agent to agent, agent to human, run to run.
Design the boundaries badly and a better model will not save you.

Scoring It

Six Places Context Leaks

Score each low / medium / high for your workflow. Then fix the top one or two, not all six.

1 · Handoffs

Context repackaged agent to agent. Raised by many sequential agents.

2 · Compression Loss

Detail dropped to make context fit. Raised by summarising between steps.

3 · Semantic Drift

Meaning mutates across paraphrases. Raised by free-text handoffs with no schema.

4 · Verification Burden

What it costs to check the output is right. Raised by opaque results with no evidence.

5 · Tokens & Latency

Raw compute of moving context. Raised by re-sending large windows every hop.

6 · Governance

Permissions, approvals, audit. Raised by unclear authority and manual sign-offs.

The Levers

What Actually Cuts It

Sink Lever
HandoffsFewer boundaries. Shared state instead of relay.
Compression lossKeep a source of truth. Summarise for display, never for the next step.
Semantic driftTyped handoff schemas. A structured contract, not prose.
Verification burdenEvidence travels with the output. Validate the trace, not the vibe.
Tokens & latencyPass references and IDs, not full payloads.
GovernanceExplicit permissions, written into the interface.
PROSE HANDOFF PROSE PROSE PROSE NOTE BRIEF DRAFT ANSWER THE SOURCE IS GONE AFTER THE FIRST HOP TYPED HANDOFF SCHEMA SCHEMA SCHEMA NOTE CLAIM + QUOTE CLAIM + QUOTE ANSWER THE ANSWER STILL POINTS BACK AT THE SOURCE

Notice that none of these are "use a bigger model."

Live · Colab

Score our own pipeline against the six

Notebook §5  ·  where the note actually went

The brief alone inflates a 372-character note ~3×, then it is squeezed back.
Nobody wrote anything wrong. The cost is at the boundaries.
Five of the six are wiring, not model. A better model pays them too.

Part II · 15 min · demos C, D

Organising the Collective

Five Shapes

How You Wire Them Matters More Than How Many

BOARD

Blackboard

Agents read and write a common evidence store.

Prefer
POLICY

Adaptive

A policy picks the structure per task.

Best

Committee

Agents argue to consensus in free text.

Worst

Hierarchy

A manager agent routes the work.

Lossy

Rigid pipeline

A fixed sequence, every time.

Brittle

The bottom three are agents imitating human meetings. That imitation is the tax.

The Worst Case

Why Debate Fails

Three agents arguing to consensus feels rigorous. It is the opposite, for two reasons.

Correlated deliberation. Same base model, same training, same blind spots. Three votes, one opinion. You have bought agreement and mistaken it for confidence.

No evidence preserved. The argument happens in prose and evaporates. What survives is a conclusion with no trace.

What To Build Instead

Give them a shared store. Each agent writes findings with provenance, reads what others wrote, and disagreement becomes a visible conflict in the data rather than a louder paragraph.

Now a human can look at the conflict and decide. That is the whole point.

Live · Colab

Same four agents, rewired

Notebook §6  ·  shared board + typed handoffs

Same answer again — but every claim now carries its source quote.
Nothing checkable → all of it, by string match. No judge model.
And cheaper than the chain. Still dearer than one agent.

The Counterintuitive One

Diversity, Not Headcount

Every time you want to add an agent, one question: is it genuinely independent?

Keep it

A different model.
A different tool.
A different evidence source.

It can be wrong in a way the others cannot. That is what adds signal.

Cut it

The same model with a rephrased prompt, wearing a different job title.

Pseudo-diversity. It adds cost on every boundary and contributes no independent signal.

Most "multi-agent systems" are one model in five hats, paying five times for one opinion.

Live · Colab

Three reviewers that are one reviewer

Notebook §7  ·  personas, then a second evidence source

Optimist, skeptic, pragmatist: same verdict, same reasoning.
They share evidence, so they share the blind spot.
One extra document finds the real cause. Diversity is evidence, not personality.

Part III · 8 min · demo E

The Boundary

The Contract

Six Things Every Boundary Must Name

Anywhere machine output enters a human workflow or another agent. Miss one and the boundary is un-checkable.

Output allowed

Which outputs may cross at all.

Evidence carried

The trace the receiver can actually check.

Uncertainty disclosed

So confidence is not laundered on the way across.

Permissions

What the producing agent was allowed to do.

Traces preserved

What is kept for audit afterwards.

Mandatory judgment

Where a human must decide — and who that human is.

Live · Colab

The handoff you actually ship

Notebook §8  ·  six fields, and a name

Output allowed · evidence · uncertainty · permissions · traces · judgment point.
Every field is mechanical. The last one is a person.
What changed? Why is it safe? What if it is wrong?

The One That Gets Skipped

An Answerable Owner

At every point where a human must decide, one named person has to be able to answer three questions about what the machine produced.

> What changed?
> Why is it safe?
> What if it's wrong?

Not the team. Not the system. A person, with a name.

Why This Is the Whole Thing

A boundary with no answerable owner is where the machine ships unowned work.

Every automation failure you will read about in your career has this shape: the output was plausible, it crossed a boundary nobody was accountable for, and by the time it was wrong it was in production.

You cannot delegate this one to the architecture.

Part IV · 5 min

What Doesn't Move

The Part That Isn't New

What Doesn't Move

Everything so far has been about systems. This is the part that was never about the tools.

Intent

You did it on purpose, and you can say why.

Curiosity

You want to know how it works, not just that it worked.

Staying True to the Work

You don't cut the corner nobody would see.

Ambiguity

You can hold a problem open before you close it.

A course can teach you principles and methods. It cannot make you any of these. Neither can the machine: it takes instruction rather than intent, and it cannot sit in ambiguity — it resolves to the average, every time.

Think For a Minute

What Did Your Course Teach You?

Two questions. The first one is uncomfortable, so be honest with yourself.

  1. What did your training teach you that a machine now does?
  2. What did it never teach you at all?

Your course gave you principles and methods, and that is the part that transfers. What it could not give you is the disposition. Unlike the methods, you build that by doing the work, which means you are not behind. You are early.

Take It With You

Four Things Worth Keeping

Cost, not capabilityThe bottleneck is what context costs to move, not how big the model is.
Wiring, not headcountShared state beats debate. One model in five hats is still one opinion.
Evidence, not outputIf the trace didn't travel with it, you can't check it, so you don't know.
A name, not a systemEvery mandatory decision needs a person who can say what changed and why it's safe.

Build the smallest system you can actually account for. Then make it bigger only when something forces you to.

Run it yourself. The notebook behind all six demos, on a free AI Studio key.

colab.research.google.com/drive/1brSW2oJTzEi4_0bd2698pYjEQVsEfJSr?usp=sharing

Then, when you want the production shape of this: Long Horizon, a reference agent harness in Google’s ADK samples (sample code, not a supported product) — github.com/google/adk-samplescore/python/long-horizon-harness

theproductguy.xyz /two-speed-engine