Research and Code Agency

An AI-driven multi-agent setup built on Agency Swarm. A manager orchestrates specialists for web research, Python execution, and interactive desktop control via E2B’s desktop sandbox to complete end-to-end tasks with clear delegation and status updates.

Agents and strengths

• ManagerAgent: Intake, planning, and delegation. Confirms goals, constraints, priorities, and risk tolerance; assigns work to the right specialist and keeps updates concise.
• WebSearchAgent: Live web research with hosted search. Pulls current facts, citations, competitor intel, pricing snapshots, news, and quick summaries with source callouts.
• CodeAgent: Runs Python in an IPython interpreter. Handles data cleaning, CSV transforms, feature engineering, analytics, validation checks, and small utilities.
• ComputerUseAgent: Operates an E2B desktop sandbox (browser and GUI). Navigates sites, clicks, types, scrolls, launches apps, and captures screenshots or stream URLs for visual feedback.

How they collaborate

• Intake → Plan: ManagerAgent confirms objectives, constraints, deadlines, and approval points.
• Research → Synthesis: WebSearchAgent gathers current evidence; ManagerAgent reviews and aligns with goals.
• Compute → Validate: CodeAgent processes data, checks quality, and summarizes results for decision points.
• Act on UI: ComputerUseAgent performs on-screen steps (form fills, uploads, confirmations) with progress notes and optional streams/screenshots.
• Loop and verify: ManagerAgent requests clarifications, re-runs steps if needed, and confirms completion criteria.

Example use cases

• Research→Code→Action pipelines: collect live info, compute metrics, then submit forms or updates in a browser.
• Competitive or market sweeps: gather sources, extract pricing/features, compute deltas, and export CSVs with highlights.
• Data wrangling and QA: clean/reshape CSVs, compute aggregates, validate distributions, and flag anomalies before handoff.
• Ops assistance: fill web forms, upload documents, capture proof via screenshots/streams, and confirm results.
• Content preparation: collect references, draft concise outputs, and place text into web editors when asked.
• Light automation with human-like steps: navigate portals, click through flows, and report any blockers or authentication prompts.

Interaction norms

• ManagerAgent is the entry point: it clarifies intent, sets plan, and routes tasks to specialists.
• WebSearchAgent cites sources and asks for clarifications on ambiguous queries.
• CodeAgent keeps snippets short, reports outputs succinctly, and explains errors with fixes.
• ComputerUseAgent narrates intended actions, avoids risky downloads, and surfaces stream URLs or screenshots when helpful.
• The system requests confirmation before high-impact actions (sending data, irreversible changes).

Inputs and outputs

• Inputs: natural language requests plus optional constraints (time ranges, targets, formats), and environment variables for keys.
• Outputs: concise summaries, citations for research, CSVs or computed metrics from CodeAgent, and screenshots/stream URLs from ComputerUseAgent runs.

Environment and credentials

• Required: OPENAI_API_KEY.
• For desktop control: E2B_API_KEY.
• Provide additional context (targets, sample data, URLs) to reduce clarification cycles.

How to run

1. Set environment variables (at minimum OPENAI_API_KEY; add E2B_API_KEY for desktop control).
2. Install dependencies: pip install -r requirements.txt.
3. Run locally: python agency.py (terminal demo) or python main.py (FastAPI entry).