All Posts
ArchitectureClaude Agent SDKAgentic AI

Engineering Blog

We Replaced Our Marketing Team with Claude Agent SDK and a Slack Channel.

Here’s the Architecture.

#marketing
Published byLightCI|April 2026

01Why Now

Two things happened in the same week that made this inevitable.

On March 31, Salesforce shipped 30+ AI features for Slack, rebuilding Slackbot from the ground up as an agentic orchestration layer with MCP client integration. A week later, on April 8, Anthropic launched Claude Managed Agents in public beta — production-grade infrastructure for running autonomous agents with sandboxed execution, persistent memory, and end-to-end tracing.

The interface layer and the intelligence layer both went production-ready at the same time. That’s not a coincidence. That’s a platform shift.

Medvi

$1.8B

projected 2026 revenue. One-person telehealth startup using ChatGPT, Claude, Midjourney, and ElevenLabs in place of an entire team.

PYMNTS, April 2026

Notable Capital

500+

BD intros managed annually with a two-person team using Claude with MCP integrations.

Notable Capital, 2026

Anthropic

70–80%

of technical employees use Claude Code daily. Internal feedback hits their Slack channel every five minutes.

Cobus Greyling, Medium 2026

We’re an AI consulting firm. If we can’t run our own marketing on agents, we have no business telling PE portfolio companies they should. So we built it.

02The Architecture

Three agents. One channel. Full content lifecycle.

LightCI's marketing function runs as a multi-agent system orchestrated through a single Slack channel. Three specialized agents — each built on the Claude Agent SDK — handle everything from ideation to published page.

System Architecture

🛠

Strategist

Research & Ideation

💬

Slack Channel

Orchestration Hub

✍️

Writer

Drafting & Narrative

🏗️

Builder

Visual & Code

Human seed → Strategist research → Writer draft → Builder ships → PR ready

🛠

The Strategist

Research & Ideation

Monitors the news cycle, identifies content opportunities anchored to real events within the last 14 days.

Produces structured briefs with source citations — no vibes, no generic "AI trends." Real events, real data, real angles.

Cross-references existing content to prevent overlap.

Never writes copy — only researches and proposes.

Web SearchSource VerificationContent Audit
✍️

The Writer

Drafting & Narrative

Takes the Strategist's brief and produces complete drafts — structured for the web, cited with real statistics.

Researches independently, pulling fresh data to fill gaps. Every stat is sourced. Every claim is backed.

Outcome-focused. Never talks about deploys, branches, or sprints. Talks about what the product does.

Never proposes topics — only drafts from approved briefs.

Deep ResearchSource CitationVoice Guidelines
🏗️

The Builder

Visual & Code

Takes the Writer's narrative and builds production React components — animations, interactive elements, responsive layout.

Every piece gets bespoke visual treatment. No templates. No markdown-to-HTML.

Registers routes, adds metadata, opens PRs. Code-ready output.

Never edits narrative — only builds from approved drafts.

ReactFramer MotionTailwindGit

03The Handoffs

A content cycle in five Slack messages.

The orchestration happens in plain English. Here's what a typical content cycle looks like.

#marketing

👤
Human9:14 AM
“Anthropic just shipped managed agents and Salesforce rebuilt Slackbot as an MCP client in the same week. I think there’s a piece about how the interface layer and intelligence layer going production-ready simultaneously changes what a ‘team’ can look like. Find the data to back it up.”

One message. A thesis, not a task. No Asana ticket. No 30-minute kickoff.

🛠
Strategist9:15 AM
Researched the last 14 days. Three ranked angles, each anchored to real events. Top pick: The convergence angle — Salesforce’s MCP integration + Anthropic’s managed agents + Medvi hitting $1.8B as a one-person company all landed in the same window. The “team as architecture” framing has hard data behind it now. 7 sources attached.
👤
Human9:17 AM
“Good angle. But don’t make it abstract — use our own system as the case study. Show the actual architecture, the actual Slack messages, the actual costs. Meta-narrative: this post was built by the system it describes. That’s the proof.”

Two human messages. A thesis and a creative direction. Everything downstream is autonomous.

✍️
Writer9:22 AM
Draft complete — 2,500 words, 7 cited sources. Includes interactive calculator spec for compound error rates, before/after flow diagrams for implicit-need detection, and builder notes on visual emphasis per section.
🏗️
Builder9:31 AM
Built and shipped. 680 lines of production React. Interactive compound error rate calculator, side-by-side flow diagrams, benchmark comparison table, three failure-recovery cards. PR ready for review. Three files touched.

2

Human messages

~17 min

Seed to PR

680

Lines of production code

04What Makes It Work

Four principles. Not the model — the system design.

Anyone can wire up three Claude calls and call it a multi-agent system. Here's what separates a demo from a functioning content operation.

1

Specialization over generality

Each agent has a constrained system prompt that defines not just what it does, but what it refuses to do. The Strategist never writes copy. The Writer never proposes topics. The Builder never edits narrative.

LangChain's State of Agent Engineering found that quality is the #1 barrier for 32% of teams. Specialization solves quality at the system level, not the prompt level.

2

The channel is the memory

Slack isn't just the interface — it's the shared context. Every agent sees the full conversation history. The Builder reads the Writer's notes. The Writer reads the Strategist's sources. No lossy handoff documents.

This mirrors what Salesforce just built: Slackbot as an MCP client that orchestrates agents through conversation context. We arrived at the same architecture from a different direction.

3

Human-in-the-loop at decision points, not execution points

The human makes exactly two decisions: "what should we write about?" and "which angle?" Everything downstream — research, drafting, visual design, component architecture, PR creation — is autonomous.

McKinsey's data supports this: 23% of enterprises are scaling agentic systems, but the ones succeeding keep humans at strategic inflection points while letting agents own execution.

4

Every agent does its own research

The Strategist researches to find angles. The Writer researches to find stats. They don't just relay information down the chain — each agent independently verifies and enriches.

Same principle behind Anthropic's "ant-fooding" culture: stress-test every layer independently, don't trust the upstream output blindly.

05What Breaks

We'd be lying if we said this runs perfectly.

The context window is the ceiling

A 200k-token context window sounds infinite until you're on the third round of revisions and the agent starts losing track of earlier feedback. Long editorial cycles degrade quality faster than fresh drafts.

Mitigation

Start fresh sessions for major revisions rather than appending to a long thread.

Agents are bad at knowing when they're wrong

The Writer will produce a stat with complete confidence that turns out to be hallucinated — a real source, a real publication, a fabricated number.

Mitigation

Verification is built into the process, but every piece requires a human review pass before publish.

Tone drift is real

Over multiple pieces, agents develop subtle stylistic tics — favorite transitions, repetitive sentence structures, predictable section patterns.

Mitigation

Voice guidelines in the system prompt help but don't eliminate drift entirely. Periodic human-authored pieces reset the baseline.

The "last mile" is still human

Publishing, distribution, social scheduling, paid amplification — the atoms-level work of getting content in front of people — is still manual.

Mitigation

The agents produce the content. They don't yet own the distribution. That's the next frontier.

06The Economics

Let's talk about what this costs versus what it replaces.

Annual content production cost

Traditional Marketing Team

$320K–$430K

Content Strategist
$90K–$120K
Senior Copywriter
$80K–$110K
Frontend Developer
$120K–$150K
Freelance/overflow
$30K–$50K

Agent-Powered Setup

$3K–$5K

Claude API
~$2K–$4K
Managed Agents
~$500–$1K
Slack
$0 incremental
Human (~3 hrs/wk)
Opp. cost only
~0%

reduction in direct content production costs

The output quality? In three weeks, we shipped a 3,000-word tactical playbook with a visual timeline and three-tier framework, an engineering deep-dive on agentic reliability with an interactive calculator, and this blog post. Each with bespoke React components, not templates.

One in three CMOs at companies over $20B expect to lay off significant marketing staff within 24 months because of AI. The question isn’t whether this shift happens. It’s whether you’re the one building the system or the one being replaced by it.

07Your Opportunity

We built this for ourselves. But the architecture is the product.

Every PE portfolio company has the same content bottleneck: strategy lives in one person's head, execution depends on a patchwork of freelancers and agencies, and the feedback loop between 'we need a case study' and 'it's live on the site' is measured in weeks.

The agent architecture we’re running isn’t specific to marketing. The pattern — specialized agents, Slack orchestration, human-in-the-loop at decisions, autonomous execution — applies to:

Portfolio Monitoring & Reporting

Analyst agent pulls data

Narrative agent writes commentary

Visualization agent builds the deck

Due Diligence Workflows

Research agent scans filings

Risk agent flags anomalies

Summary agent produces the memo

Customer Operations

Triage agent classifies tickets

Resolution agent drafts responses

Escalation agent routes to humans

0%

of enterprises are experimenting with AI agents

0%

are actually scaling them

McKinsey, Global AI Survey 2025

The gap is the same one we closed: moving from “we use AI tools” to “AI tools are the team.”

The firms that figure this out during the hold period don’t just improve margins. They walk into diligence with an operating model that buyers want to acquire — because the agents come with the company.

91% of marketing teams use AI to help humans do their jobs.

We used it to redesign what the job is.

 

Three agents. One Slack channel.

A human who points and decides.

 

Everything else is architecture.

Yes, this blog post was written and built by the system it describes.

Sources & References

[1]

Anthropic, "Claude Managed Agents" launch — SiliconANGLE, April 8, 2026

[2]

Salesforce, "30+ AI Features for Slack" — TechCrunch, March 31, 2026

[3]

PYMNTS, "The One-Person Billion-Dollar Company Is Here" — April 2026

[4]

Intuition Machine, "Antfooding: How Anthropic Builds Products" — Medium, 2026

[5]

MarTech, "The Truth About Martech in 2026" — 91% AI adoption

[6]

LangChain, "State of Agent Engineering 2026"

[7]

McKinsey, "The State of AI: Global Survey 2025" — 23% agentic scaling

[8]

Anthropic, "Building Agents with the Claude Agent SDK"

[9]

Tech.co, "Companies Replacing Workers with AI in 2025–2026"

Ready to move

Ready to build something like this?

We design and deploy AI-native systems for companies moving fast in competitive markets.

Talk to LightCI