how to make your company AI-ready
your agents are only as good as what they can read, so making a company AI-ready comes down to the setup underneath them, which are three layers
> company brain, all your raw company data centralized and turned into a wiki your agents can read
> data warehouse, the numbers side of that brain, everything you track
> the operating system, the agents that read the brain and run the work
the brain has a simple shape:
> sources, what was said, your calls, chats, docs and transcripts
> data, what you can measure, your metrics, exports and tables
> wiki, what you know, sources and data compiled into linked pages
sources and data both feed the wiki, that's the layer your agents query
there are more complex ways to build this, but the simplest reachable version is a karpathy-style LLM wiki in one folder:
> a sources folder, markdown files for chats, transcripts and docs
> a data folder, a sqlite database for your datasets, keeping the warehouse simple
> the wiki, an agent compiles sources and data into cross-linked pages
> semantic search on top so it finds by meaning even when the words don't match
you can move the warehouse to postgres when it outgrows sqlite, for now the whole thing is one file next to the wiki
then the operating system sits on top, a layer you and your agents work the brain through, where the day to day work gets done
> an orchestrator routes each job and hands the right agent its context
> a specialist agent, or a small team, owns each vertical and reads the brain before it acts
> you steer the direction, review the output, and approve what ships
> every result files back into the brain, so it compounds
keep the surface simple to start, a chat or the wiki wired to notion or drive is enough, the orchestrated agent company comes later
Loop engineering - the reading list
In 2026 agents stopped being about smarter prompts and started being about longer runs.
Everyone needs to stop writing stupid prompts and start learning Loop engineering.
The real question isn't "what do i type". It's "how does my agent keep going for 40 minutes without falling over".
1. Can it recover from a failed step?
2. Can it control spend?
3. Does it know when to stop?
All of it comes back to loop design:
[ READING LIST ]
1. Addy Osmani - Loop engineering:
https://t.co/kzIbYW8wLG
2. Firecrawl - Loop engineering:
https://t.co/8UhKcZvbw9
3. Oracle - What is the AI agent loop:
https://t.co/Jg5ic7dxJc
4. OpenAI - Harness engineering:
https://t.co/7i34jS1Qk9
5. Martin Fowler - Harness engineering for coding agent users: https://t.co/1QvsIHGbXa
6. From React to loop engineering - Agentic loops:
https://t.co/WERkgRXWsy
7. Mem0 - Loop engineering for ai agents, memory-first:
https://t.co/mJxzguwX7z
[ OPEN SOURCE WORTH READING ]
1. Codex CLI: https://t.co/TCbo5tNb3b
2. Openhands: https://t.co/KgPJOHgLK4
3. Pydanticai: https://t.co/6Dd1Hu9Etj
4. OpenAI Agents SDK: https://t.co/JobwcV75dH
[ WHAT TO STUDY ]
- How the loop runs?
- How the loop stops?
- How the loop verifies?
- How the loop recovers?
- How the loop is debugged?
[ THE POINT ]
- Prompt decides how the agent starts.
- Context decides what the agent sees.
- Loop decides how far the agent gets.
Scheme:
Think -> Act -> Observe -> Verify -> Evolve -> Repeat
[ START HERE ]
Before you touch anything above - read my Article first -It's the entry point.
WorkBuddy + ima 是小白搭建个人知识库最容易上手的组合:
ima 负责存资料(微信、PDF、会议、视频),WorkBuddy 负责读取整理、摘要和生成 review,你自己把关最终判断。
第一版只建3个库(01-资料库、02-项目库、03-输出库),先少量测试跑通“资料进来后有人处理”这一步,避免变成第二垃圾桶。
核心是养成每周复盘习惯,让资料真正复用,而不是越存越乱。