Framework for autonomous agents that remember and search

What it does

OpenAGI is a Python framework for assembling autonomous agents from interchangeable parts. You define Worker roles, assign them tools like DuckDuckGoSearch or TavilyWebSearchQA, and let an Admin agent orchestrate the plan. It supports multiple LLM backends—currently OpenAI and Gemini—and includes a long-term memory module so agents can reference past interactions across sessions.

The interesting bit

The framework treats human intervention as a dial, not an afterthought: the TaskPlanner exposes a human_intervene flag so you can choose whether the loop stops for a human check. That is a rare bit of UI-level thinking in a space that usually assumes full autonomy or full manual mode.

Key highlights

• Pluggable LLM support via OpenAIModel and GeminiModel classes.
• Built-in tool integrations for DuckDuckGo and Tavily web search.
• Long-term memory module (Memory(long_term=True)) for recalling previous queries and context.
• Two execution modes: explicit multi-worker teams or leaner autonomous Admin runs.
• Task decomposition handled by a dedicated TaskPlanner with optional human-in-the-loop gating.

Caveats

• The README is heavy on aspiration—“AGI for all”—and light on architectural detail; it is unclear how memory is persisted or how workers coordinate under the hood.
• Examples show API keys set inline with os.environ, which suggests the project is still in a “script-first” stage rather than production-hardened configuration.
• The project does not explain what differentiates it from similar agent frameworks like CrewAI or AutoGen.

Verdict

Worth a look if you want a quick, Pythonic way to prototype multi-agent LLM workflows with search and memory. Pass if you need mature observability, deep architectural docs, or enterprise-grade security patterns.