Episode 35: Agent Inspection UI
We review our nearly-done inspection UI and run our GitHub agent, seeing each step's input and output showing cleanly in the UI.
Next we'll deploy to production and invite a few people to try it out!
Episode 34: Agent Inspection UX Design
We review our basic inspection UI and the refactored code that now stores each agent action to the database. We reopen Excalidraw and design how the dashboard will display all task details.
See the diagram 👀👉 https://t.co/PAqHGCx70t
Episode 33: Agent Inspectability Planning
We think about agent inspectability from first principles and spec out a public web dashboard enabling anyone to inspect the input/output and metadata for every step taken by an agent.
Read the spec 👀👉 https://t.co/9BpysDyn5Z
Episode 32: Toward Semi-Automation
We review Faerie’s pull request which includes multiple rounds of commits factoring in user feedback and automated test results.
Only a bit of prompt engineering stands between us and a mergeable PR. Next we'll add visibility into each step of the agent’s thought process so we can more easily tweak relevant prompts.
👀👉 https://t.co/LQIPivOmgo
Episode 30: Faerie Debugs Failing Tests
We modify our GitHub Action to add a comment to the PR detailing the failing tests, then Faerie adds a comment with a fix. Next she'll commit the fix so failing tests will pass 😎
👀👉https://t.co/mm5keufMiO
Episode 28: Creating New Files
We give Faerie the ability to create new files for her pull request. The quality of the PR improves dramatically and the code looks almost ready to merge. Next we verify the code quality with tests!
👀👉 https://t.co/uxc0JAtFiX
Episode 27: Smarter Pull Requests
We tweak our prompts to add more and better context to Faerie's PRs. She writes multiple simultaneous commits that improve on our code, but struggles because we haven't let her create new files. That's next!
👀👉 https://t.co/5lF5Y6otRK
Congrats to ClosedAI (dba "@OpenAI") on the bailout from your daddy Microsoft!
We’re excited your puppet strings are on full display — and that you’ll be at full strength for what’s about to happen.
See ya soon 😘
Episode 26: Faerie Commits Code
Faerie makes multiple code commits to a GitHub branch based on an issue, then opens a pull request.
After another day or two of prompt engineering, her PRs should be mergeable. 🎉😎
👀👉 https://t.co/iUpjRTgr39
Microsoft is the new end boss.
They just acqui-hired the OpenAI core team.
As we said 2 weeks ago:
"We don't think the future of AI should be owned by closed-source megacorps with a history of monopolization and regulatory capture.
Now we organize a true counterweight.
FRAME IT 👇
Between web & workerbee (https://t.co/KMiQtqLKyD) we have about ~200 GPUs connected - which wethinks makes us one of the largest decentralized clusters already - and that should 10-100x soon after OpenAgents goes live & builds demand
Truly open AI at scale⚡️🪄✨🎉
Episode 25: Faerie Writes Code
Faerie completes the first step of her plan by writing code for the Memory model, database migration, and unit tests.
After two quick fixes from a human, the tests pass and our first AI-generated code is merged to the OpenAgents codebase. 🎉
The code 👉 https://t.co/Zy3lCWeCne
Episode 24: Faerie Makes a Plan
We give our agent all relevant code via retrieval from our embeddings and ask her to make a plan to solve the issue. She replies with a brilliant plan detailing specific changes to our code, which she'll next begin to write!
Read the plan 👉 https://t.co/I4YL6Vi1sk
Episode 22: Conversing with Faerie
We add all the issue comments as context to the prompt, enabling our new agent to give smart reflections on the entire conversation.
Faerie's second comment 👉https://t.co/UFTLBwUlLQ