Fivetran published a blog post this week titled "Anthropic, Please Make a New Slack." It went viral for good reason. The frustration is real. Slack was built for humans typing messages to other humans. It was never designed for a world where half the participants in a channel might be AI agents.
And Slack isn't the only broken tool. It's all of them.
The Problem With Every Tool You Use
Your CRM was designed for humans clicking through forms. Your project management tool was designed for humans dragging cards across boards. Your email client was designed for humans reading and replying to messages one at a time.
None of these tools have proper APIs for agent-driven workflows. Or if they do, the APIs are afterthoughts: rate-limited, poorly documented, missing half the functionality of the UI.
Fivetran's frustration with Slack is the same frustration every company hits when they try to connect AI agents to their existing stack. The tools fight back. Slack's API has bizarre rate limits. Google Calendar's API requires an OAuth dance that would make a contortionist jealous. CRM APIs return data in formats that require 200 lines of parsing code.
Why "Just Use the API" Doesn't Cut It
People who've never deployed an AI agent always say the same thing: "Just connect it to the API." As if every API is a clean, well-documented, reliable interface.
Reality check: Slack's rate limit for posting messages is 1 per second per channel. If your agent needs to triage 50 messages across 10 channels, that's already a bottleneck. Want to search message history? Different rate limit. Want to react to messages? Different rate limit. Want to update a message you already sent? Believe it or not, different rate limit.
Now multiply that by every tool in your stack. Each one has its own auth flow, its own rate limits, its own data formats, its own quirks. Connecting an AI agent to 8 tools means solving 8 completely different integration puzzles.
This is exactly the problem we solve at OpenClaw Setup. We've already fought these battles. We know which APIs are reliable, which ones need workarounds, which ones require specific auth configurations, and which ones will break if you look at them wrong. That knowledge is baked into our deployment process.
The Real Question: Should We Wait for Better Tools?
Fivetran wants Anthropic to build a new Slack. That's one approach. Wait for someone to build tools designed for AI from the ground up.
Here's my take: don't wait.
New tools take years to build, years to reach feature parity, and years to get adopted. Slack isn't going anywhere. Neither is Google Workspace, or Notion, or HubSpot. Your business runs on these tools today.
The smart play is to build an agent layer that works with your existing tools, messy APIs and all. When better tools eventually arrive, your agent adapts. The agent is the constant. The tools underneath it are variables.
What Actually Works
After deploying agents for dozens of businesses, here's what we've learned: the integration layer is 70% of the work. The AI model is maybe 15%. The prompting and behavior design is the other 15%.
Most companies that try to build their own agent get the model working in a weekend and then spend 4 months fighting APIs. That's backwards. You need a team that's already solved the integration problems.
That's literally what we do. $999, your agent is deployed, connected to your tools, and running in your Slack. The API headaches are our problem, not yours. Let's talk.