lunch & learn

Behavior-Centric
AI Design

Nick Stinemates
Co-Founder, GTM
Swamp Club
who we are

Builders

Technologists and AI people first

The team

Adam Jacob

CEO
Prior
Chef::HJK

Mahir Lupinacci

COO
Prior
Chef::SourceClear

Nick Stinemates

Co-Founder, GTM
Prior
Docker::Kentik

Paul Stack

VP Product
Prior
HashiCorp::Pulumi

John Watson

Head of Platform
Prior
Matillion::Deloitte
how it started

Swamp was a codename that became the company

Two ideas

  • 01Use AI in every job in the business
  • 02Build software for YOU

Swamp runs Swamp

Engineering
Finance
Forecasting
Sales
Customer success
This deck

A bug you file

~3 hrs
until someone looks at it
+1 hr
until it’s fixed

Everything in this talk, we did to ourselves first

a fundamental shift

From early experimentation to at-scale production use

where we started

Agent-centric

The model decides the control flow

where we're going

Behavior-centric

Frontier intelligence only when the job requires it

agent-centric ai design

The loop

The LLM is the heart of the program

The agent loop

Monolithic agents

  • 01Grow it: vertical scalability problems
  • 02Split it: a distributed monolith

Plan a trip to Japan

The more an LLM is responsible for the control flow, the more correctness relies on it making the right decision at every point

Rework and validation

  • 01Double check the itinerary
  • 02Try the hotel agent again
  • 03Add an agent to validate the agent
  • 04Run the entire loop again
to the person who built it

A papercut

to everyone else

A business trip without a hotel

The ROI death spiral

40%+
of agentic AI projects will be cancelled by 2027
Gartner
63%
of enterprises using AI can’t show any financial impact
McKinsey
the next layer

Skills

Context and scripts, bundled for the model

a hundred times a day

Instead of going to the LLM every time… let’s just record that and turn it into deterministic scripts

Why skills alone fail

The model still decides
Tokens on every run
Nothing is remembered
Nothing to audit
Sharing is unsolved
Sometimes it just doesn’t run

The goal is the business outcome

outcomeThe job gets done, every time
costKnown, and it doesn’t balloon at scale
safetyControl, audit, compliance
ai outcomes

AI ability is unevenly distributed across teams

The Pareto principle

20%
of your people
80%
of the AI results
behavior-centric ai design

Behaviors

A corpus of reusable, composable building blocks

Every behavior

Typed, validated inputs and outputs
Data in versioned schemas
Indexed for search
Its own credentials

Skills are direction

skillDirection
behaviorsAgency to reach into a system
workflowsThe prescriptive thing, every time
plan trips

Decompose your intent

  • 01Haiku.extract()
  • 02Flight.search()
  • 03Hotel.search()
  • 04Itinerary.pick() + generate()

Each LLM call is scoped to a single, well-defined job and never drives the control flow

Regular old deterministic code

  • 01No tokens spent on control flow
  • 02No validation after the first step
  • 03The same shape every time
  • 04A fixed and understood cost

The trip planner, both ways

14x
fewer tokens
total tokens per run
6x
cheaper to run each plan
model cost per execution
turn it into an agent

One call

The agent should decide what to do. The system should know how to do it.

Use frontier intelligence at build time to encode our intent into behaviors we can rely on

what your teams already wrote

The Python in your skills is a behavior waiting to happen

From a script to a shared behavior

  • 01Claude turns the script into a behavior
  • 02It lives in source control, reviewed like code
  • 03It’s published to your private collective
  • 04The skill points at it

The skill ships with the behavior

manifest.yaml manifestVersion: 1 name: "@yourco/brand-assets" version: "2026.11.19.1" visibility: private behaviors: - brand_assets.ts workflows: - publish_campaign.yaml skills: - brand-assets

Built once, used everywhere

marketingBuilds the behavior
videoRuns it in their workflow
accountingRuns it in theirs

Vetted, managed, shared

Private to your company
Reviewed in source control
Promoted from beta to stable
Each repo picks who it trusts
Dependencies audited at publish
Credentials the agent can’t see
shadow it

Sanctioned

The same work, in the open, under guardrails you set

The software factory

  • 01Every review and validation step is a behavior
  • 02The factory is a workflow
  • 03It runs the same for every person who uses it

Code review, both ways

8x
fewer tokens
production code review workflow
100s
of ships a week, per developer
software factories on swamp
swamp is

The equalizer

The top 20% build the foundation for non-technical teams to thrive with their work

xero

Our marketers know exactly who they want to reach, and they shouldn’t need a DevOps engineer to act on it

Adam Roberts · SVP Data Science, Analytics & Insights

Swamp Club

340M+
runs
1,500+
extensions
100+
write-ups

Get started

~$ curl -fsSL https://swamp-club.com/install.sh | sh ~$ swamp start
no commands to learn

swamp start

  • 01Sign in
  • 02Pick or create a workspace
  • 03Describe the task in plain words
  • 04Pick your agent
  • 05It opens, already working on it

Start with what you repeat

>Look at my agent session history and find repeatable tasks I can encode as swamp behaviors.

Then

  • 01Compose behaviors into a workflow
  • 02Publish to your collective, privately
  • 03Hand it to a team that isn’t yours
thank you

It’s time
to build!

Nick Stinemates
Co-Founder, GTM
Swamp Club QR code for https://swamp-club.com/campaign/nick-stinemates/qr