Overarching goal: Focus tech development efforts around making data, models, and agents accessible to every employee – not traditional software development teams building app interfaces.
Traditional software developers have a longstanding disconnect with their users: They aren’t the ones using the app. As a result, many apps are developed that users won’t use, don’t like to use, or doesn’t fit their needs, and adoption struggles as a result.
With the advent of AI agents that excel at writing code and manipulating data at 1000x the speed of a human, this all changes. No longer are traditional software developers needed to develop apps and users no longer need to offload the work of developing apps to another human. Instead they can build interfaces to their data quickly themselves. And every day, agents get cheaper, faster, and better at doing this too.
Models/Agents have the ability and speed to act on behalf of and develop the interfaces that people want to interact with their data. We should not be wasting our time trying to do that for them.
—
The software engineers of today should be using their skills to perform data engineering and organizing and advising on the structure of data we already have. And building the larger/more technical tools/MCP servers that we don’t already have.
Eventually AI will get better than them at doing this too though.
Priorities:
- Make all data (all project files, data models, platforms, teams chats, emails, etc) accessible by authenticated API / MCP servers to every employee to both read and write data. Agents will be able to use these access points to execute requests on behalf of the user.
- Establish a centralized MCP, Skills, Tools, and Knowledge server to enable agents used by employees to use and write tools, knowledge, and data to the server for their teammate’s agents’ to be able to leverage.
- Create a central model / agent api for tracking all model use company-wide
- Build systems to protect all users on model risks like prompt injection and mistakes
- Establish safeguards that test humans that they are reviewing outputs and review outputs with review agents as well.
- Give local accessible agent runtimes (opencode, claude, etc.) to every employee. Set a budget for employees and tie usage to projects.
- Optimize
- Begin organizing and centralizing all data, minimizing agent search and execution time.
- Establish tools to test different models/harnesses against each other for tasks.
- Eliminate human review where we are confident in doing so
- Give employees cloud-based agent platform / VM workspace. Agents on this platform will reside in the same datacenter in which the data resides, allowing for faster execution of work
- Gather company knowledge. Automate extracting knowledge from teams chats, emails, and existing chats users have with agents.
BUILD FROM THE TOP DOWN
Centralize data around a common database.
- Start moving all data to that
- develop a cad platform to integrate
- Stop creating files / unstructured / unlinked data