agents

AI Agents

This folder contains 2 scripts which leverage the Claude Agents SDK to automatically populate the data of the AI Deadlines web app.

Usage

First create a keys.env file at the root of the repository which contains the following environment variables:

ANTHROPIC_API_KEY=
GITHUB_PAT=
EXA_API_KEY=

In case you want to leverage MiniMax 2.1 instead of Claude, get your API key from https://platform.minimax.io/ and add the following:

ANTHROPIC_BASE_URL=https://api.minimax.io/anthropic

Next, the agent can be run like so on a conference of choice:

uv run --env-file keys.env -m agents.agent --conference_name neurips

The agent will automatically fetch relevant information from the web using the Exa MCP server to populate the data at src/data/conferences and open a pull request.

Modal deployment

To automatically let the AI agents populate deadlines data, we leverage Modal's serverless infrastructure.

Setup

1. Install Modal: uv add modal
2. Authenticate: uv run modal setup
3. Create the required secrets:

uv run modal secret create anthropic ANTHROPIC_API_KEY=<your-api-key>
uv run modal secret create github-token GH_TOKEN=<token-with-repo-and-pr-scope>
uv run modal secret create exa EXA_API_KEY=<your-key>

Note: The GH_TOKEN needs the repo scope (for cloning, pushing, and creating pull requests).

Running a single conference

uv run modal run agents/modal_agent.py --conference-name neurips

Running all conferences in parallel

By default (no flags), the script processes all conferences in parallel — each in its own Modal container:

uv run modal run agents/modal_agent.py

You can also be explicit:

uv run modal run agents/modal_agent.py --all-conferences

To test with a limited number of conferences, use the --limit flag:

uv run modal run agents/modal_agent.py --limit 3

Deploying for scheduled runs

To deploy the agent so it runs automatically every week (Sunday at midnight UTC):

uv run modal deploy agents/modal_agent.py

The structure looks as follows:

flowchart LR
    A1["Retrieval Agent 1"] --> V["Aggregation Agent\n(majority vote)"]
    A2["Retrieval Agent 2"] --> V
    A3["Retrieval Agent 3"] --> V
    V -->|"selected result"| B["Verification Agent"]
    B -->|"structured output: reasoning + verdict"| C{verdict?}
    C -->|true| D["PR Agent"]
    C -->|false| E["Return: no PR"]
    D -->|"structured output: PR result"| F["Return result"]

Loading