Currently this can block TTL for a long time

• ris’ idea: social connector like slack or gmail, you have inboxes for sharing projects or ideas for friend or colleague to give feedback or help edit or just shoot the shit - chat function
• Goal enrichment/flow at startup.
• Summary (memory.context) in GUI
• Mac or electron app, keybinding, asks you to drag and drop relevant documents / write relevant information, lil bee swarm overlays
• Provisional results (best guess at each step)
• https://codepen.io/Grsmto/pen/ARpmyg
• Railway cloud UI https://railway.app/project/f3a1467a-9621-49a5-a082-817a68bf5258 add agents, documents, etc
• I like this app:https://app.haikei.app/
• https://arxiv.org/abs/2304.11477 PDDL / LLM+P
• Frickin waggling bees man
• Web workers for chain machine
• Move chainmachine to API/backend
• Post public answers for SEO and munny
• LEARN to optimize the plan->execute->review execution loop
TWIML podcast 5/2/23 @25:00
• Heuristic re-scheduling / review
• I like the contributing section https://github.com/node-formidable/formidable
• Databerry’s docs generator Powered by Mintlify
• Prompt Selectors in langchain
• Use existing web crawler
• llama index guy says that agents that pick tools are a frontier
• https://abacus.ai/

AI-Engineer-Foundation/agent-protocol#74
https://github.com/langchain-ai/langchainjs/blob/main/.github/contributing/INTEGRATIONS.md

test implementation ./run.sh serve in autogpt repo

after the DAG from plan->review completes, we have a list of tasks, as well as a review. we must start executing these tasks.

• Consider changing the output of /plan API to a dependency graph.
• Also, possibly every agent's output.
• Concurrently, review of the plan must happen.
• Stub out means for invalidating data and the progress from chains that fail review

implementing the global store. do global stores have any downsides? testability?

CTRL+F/CMD+F "TODO:" + "FIXME:"

Currently docker support is a lie. Implement a repeatable docker build.

Currently, with the new persisted goal, it is easy to lose place in the app.

• add a goal (or waggle? what's the right name for this?) history
• add some UI affordance (like browser tabs) that hints at multiple goals being a thing.
• show the goal title/summary in the tab or whatever UI (add a Title summary prompt like ChatGPT does for convos)

Or otherwise use security/scope features of the vector db

Go through a few iterations that help make sure the goal is clear and the system is set up for success. This should happen before or around the same time as uploading documents.

Try another model?

client: success state,

message: That model is currently overloaded with other requests. You can retry your request, or contact us through our help center at help.openai.com if the error persists. (Please include the request ID ba26ecc19b891d51c802fd1bf379d34a in your message.) type: server_error param: null code: null

...server

@acme/nextjs:dev: handleChainError {
@acme/nextjs:dev: type: 'handleChainError',
@acme/nextjs:dev: err: 'message: That model is currently overloaded with other requests. You can retry\n' +
@acme/nextjs:dev: ' your request, or contact us through our help center at help.openai.com if the\n' +
@acme/nextjs:dev: ' error persists. (Please include the request ID\n' +
@acme/nextjs:dev: ' ba26ecc19b891d51c802fd1bf379d34a in your message.)\n' +
@acme/nextjs:dev: 'type: server_error\n' +
@acme/nextjs:dev: 'param: null\n' +
@acme/nextjs:dev: 'code: null\n'
@acme/nextjs:dev: }
@acme/nextjs:dev: createModel: gpt-3.5-turbo-16k

Look at similar projects, have a quick discussion and set up pull request gating.
#8

• Add the newer .env.example to env.mjs
• DX: add some kind of build error or helper script to make syncing .env.example and env.mjs painless in the future
• BONUS -- set up quick README.md for contribution pattern

• https://www.agentprotocol.ai/protocol https://github.com/AutoPackAI/Auto-GPT-Benchmarks
• https://github.com/THUDM/AgentBench
• others?

I have also taken on an initiative to add an Agent Protocol-calling Agent within Langchain, and will likely try to merge that in and use it within waggledance.ai, or vice-versa
AI-Engineer-Foundation/agent-protocol#74

Cloud infrastructure is often blocked from web scraping/using a headless browser.

Implementation ideas:

• We could introduce a new AgentPacket or Tool, requestClientExecution
• alternatively, maybe we could make a WebSocket or WebRTC connection to the client to reduce latency or "punch a hole" through to their network to use as a http proxy

Current contenders:

• upstash redis (vercel KV store)
• vector

https://openai.com/policies/usage-policies

Social engineering tiem

• react-force-graph seems to be adequate for visualizing the DAG.
• Consider making the visualization the primary feedback.
• Add some kind of node-visualization when tapping/clicking. show state of that node chain I guess?
• Tabs may not be the right interface for debugging/results/visualizing.
• Optional: https://github.com/openai/shap-e come up with 3d visualizations for each task
• Optional: Summary (memory.context) in GUI

https://github.com/spcl/graph-of-thoughts
https://arxiv.org/pdf/2308.09687.pdf

Rate limits

OpenAI enforces rate limits on the requests you can make to the API. These are applied over requests-per-minute, tokens-per-minute, or in the case of image models, images-per-minute.

Learn more in the rate limits documentation, or reference the default rate limits for our models. Your organization's rate limits are listed below.

https://js.langchain.com/docs/modules/model_io/models/chat/how_to/llm_caching

adding some local persistence layer that synchs with the prisma backend. I think it makes sense to create a goal model and start to build relations from there.

• choose documentation framework
  I like (not a constraint):
  Powered by Mintlify
  docusaurus

• I really like mendable AI doc search for cool factor and hype. I also have been using it at js.langchain.org
  https://www.mendable.ai/pricing they have a free plan for open source

• 1:
  • id: 0
    name: 📚 Research successful digital marketing startups
    act: Research
    context: ...
    successors: [1, c]
  • id: 1
    name: 📚 Study target audience and market trends
    act: Study
    context: ...
    successors: [c]
  • id: c
    name: 🔍 Review the research findings
    act: Review
    context: ...
• 2:
  • id: 0
    name: 📝 Create business plan
    act: Create
    context: ...
    successors: [1, c]
  • id: 1
    name: 📝 Set up legal and financial structure
    act: Set up
    context: ...
    successors: [c]
  • id: c
    name: 🔍 Review the business plan and strategies
    act: Review
    context: ...

is less tokens and also shows relationships sooner than current format

On /waggle-dance page:

The reason this is an x-large ticket, is because it requires potentially generating or interacting with a TON of possible outputs.

What is worth reporting? This is context-dependent, I think.
Should we focus on a few narrower use cases (like generating text files, committing to .git repositories?)

• Add a UI for downloading or otherwise receiving the results, as the chains report in. Output will likely come from a central planning agent, or from significant sub-tasks.
• Attach significant results to the waggle dance object in the model layer

https://nextjs.org/docs/app/building-your-application/upgrading/app-router-migration

The app router was still in Beta, and sometimes lacks features, while at the same time allows more (isomorphic) code sharing for APIs.

Try adopting App router for Next.js pages APIs.

After type errors and workflows are fixed, check this box in https://github.com/agi-merge/waggle-dance/settings/branch_protection_rules/new?branch_name=main

#4

Previously, I was trying middleware, but buried in the docs is the fact that it doesn't work for database adapters (e.g. Prisma.)

use jwt or manually guard api and frontend without middleware

In the documents branch, I needed a user id for the upload buckets, so I figured out and started fixing this.

• implement custom signin page
• guard api and page routes as needed
• offline-first/anon users/more graceful transition

some env vars are of known length.
we can catch a whole class of bugs with guarding for proper lengths.

placeholder/search for this:

    // Add `.min(1) on ID and SECRET if you want to make sure they're not empty
    DISCORD_ID: z.string(),
    DISCORD_SECRET: z.string(),

HTTP streaming. this would be useful for keeping the UI up to date during plan, execute, review, and to show file upload and processing status

https://vercel.com/docs/concepts/functions/edge-functions/streaming

I added a linear progress, now to just use elapsed time + whatever algorithm to estimate

My latest thinking is that it would be great to pass in some simple objects describing state, entities, etc.

this could be implemented as a tool, or in a custom AgentExecutor

possible dupe of https://github.com/orgs/agi-merge/projects/1/views/1?pane=issue&itemId=36323324

There are many type errors throughout the app.

• Fix them lol. only in extreme cases disable rules.
• Make sure ALL of the current and future contributors know how to auto-fix. I'm talking about me, Jon. I still don't get it lol. Write some docs!!!
• Strongly encouraged: Add requirement to PRs (builds currently fail, can require stricter PR requirements and Vercel requirements)

Start here: https://arxiv.org/abs/2304.11477
https://github.com/Cranial-XIX/llm-pddl

the output of this ticket should be either rejecting PDDL with cause or another ticket proposing how to integrate it.

StructuredTool abstract base class can only be used for OpenAI function calling executors. Additionally, Tool can only be used for non-structured OpenAI executors.

This can be resolved by dynamically switching between the two abstract types, based on which executor is currently being used.
It could also be resolved by making (a) better executor(s) that can all use structured tools and automatically include schema.

See: agents/src/skills/saveMemory.ts (or long-term memory)

• AbortController for fetches
• https://js.langchain.com/docs/modules/agents/additional_functionality#cancelling-requests