Agents in Production.
— SAHAR CARMEL · BUILDERS CONF 2026 —
Issue 01
Folio · I
// CONTENTS
Sahar
Carmel.
Director of AI Enablement at Mixtiles.
In this talk
i. Hello, briefly. p. 01
ii. Human-Driven Development. p. 02
iii. Claude Code → SDK. p. 03
iv. Agents, Together. p. 04
Day Job Mixtiles AI enablement across product, ops, growth. mixtiles.com
Community Squid Club A community for engineers building with AI. squid-club.com
Side Project Candlekeep Context retrieval for what models don't know. getcandlekeep.com
Agents in Production.
— METHOD · HOW TO CHOOSE WHAT TO BUILD —
Issue 01
Folio · II
// CHAPTER II.
Human-Driven
Development.
— observe, then mimic. The brief is in the seat next to you, not in the roadmap.
Agents in Production.
— OBSERVATION · THE PAIR I WATCHED —
Issue 01
Folio · III
// CHAPTER II. (cont.)
Two people,
two agents.
— I watched them work together. Then I built agents that do the same thing.
The pair
01 The analyst. Pulls data. Finds the signal.
02 The copywriter. Takes the signal. Writes the story.
→ The handoff. Data flows from one to the other.
Agents in Production.
— PIPELINE · FROM WORKBENCH TO PRODUCTION —
Issue 01
Folio · IV
// CHAPTER III.
Claude Code
→ SDK.
— they speak the same language because they are the same language.
From skill to shipped
01 Author in Claude Code. See the thinking. Iterate.
02 Lift into the SDK. Same skills. Same tape.
03 Run where work happens. Slack. Cron. Ticket queue.
Agents in Production.
— LIVE DEMO · INTERLUDE 01 —
Issue 01
Demo · I
// DEMO  ·  LIVE

Building the agent.

— plan it with context. Author it in Claude Code. Ship it to Slack.
demo · plan → claude-code → slack Live
Agents in Production.
— CRAFT · WHAT MODELS DON'T KNOW —
Issue 01
Folio · VI
// CHAPTER IV.
Teach the Model.
— same agent. Same task. Five different prompts.
Opus 4
Sonnet 4
Haiku 4
GPT-4.1
o3
Role
system
system
system
system
developer
Structure
XML tags
XML tags
XML tags
Markdown H1–H4
minimal text
"Think step-by-step"
omit
✓ add it
omit
✓ add it
✗ HURTS perf.
"MUST" / "ALWAYS"
soften ↓
✓ keep
✓ be explicit
repeat at end
omit
Prompt size
~3K tokens
~2.5K
~500
~3K (×2)
~500
Sources: Anthropic Prompting Best Practices · OpenAI GPT-4.1 Prompting Guide · Optimized via CandleKeep
Agents in Production.
— PROTOCOL · HOW AGENTS TALK —
Issue 01
Folio · VII
// CHAPTER IV. (cont.)
Agents,
Together.
— a single agent is a tool. Two agents with a protocol are a team.
Agents in Production.
— LIVE DEMO · INTERLUDE 02 —
Issue 01
Demo · II
// DEMO  ·  LIVE

The Duo.

— two agents, one channel. Watch them work.
demo · #marketing-agents · slack Live
Agents in Production.
— TAKEAWAYS · WHAT TO REMEMBER —
Issue 01
Folio · IX
Bottom Lines.
01 Watch before you build. HDD — the spec is in the seat next to you.
03 Workbench → Factory. Author in Claude Code. Ship via SDK. Same language.
02 One person, one agent. Decompose by role, not by task.
04 Teach each model its dialect. Same agent, different prompt. Read the docs — or let a tool read them for you.
05 Fit the workflow people already have. Same channel, same @mentions — adoption follows familiarity.
Agents in Production.
— SAHAR CARMEL · BUILDERS CONF 2026 —
Issue 01
Coda
// FIN
Thank
you.
— observe, build, ship, connect.
Find me
LinkedIn /in/saharcarmel
Twitter @SaharCarml
Community squid-club.com
Tool getcandlekeep.com
Day Job Mixtiles AI enablement across product, ops, growth. mixtiles.com
Community Squid Club A community for engineers building with AI. squid-club.com
Side Project Candlekeep Context retrieval for what models don't know. getcandlekeep.com