# Fringe & Experimental Frameworks

Cutting-edge, research-oriented, and niche agent frameworks worth tracking.

---

## MetaGPT

**Software company simulation** — Multi-agent with SOP (Standard Operating Procedures).

```python
from metagpt.roles import ProductManager, Architect, Engineer, QA
from metagpt.team import Team

team = Team()
team.hire([
    ProductManager(),
    Architect(),
    Engineer(),
    QA()
])

await team.run("Build a snake game")
```

| Feature | Detail |
|---|---|
| **Paradigm** | Role-based, SOP-driven |
| **Output** | Complete codebase (docs, code, tests) |
| **Communication** | Structured messages, publish/subscribe |
| **Best For** | Greenfield software projects, prototyping |
| **License** | MIT |
| **Status** | Active, production use cases |

---

## CAMEL (Communicative Agents for "Mind" Exploration)

**Role-playing framework** — Two agents collaborate via structured dialogue.

```python
from camel.agents import ChatAgent
from camel.messages import BaseMessage

assistant = ChatAgent(
    system_message=BaseMessage.make_assistant_message(
        role_name="Programmer",
        content="You are a Python programmer."
    )
)

user = ChatAgent(
    system_message=BaseMessage.make_user_message(
        role_name="Client",
        content="You need a web scraper."
    )
)

# Role-play
for _ in range(10):
    user_msg = user.step(assistant_msg)
    assistant_msg = assistant.step(user_msg)
```

| Feature | Detail |
|---|---|
| **Paradigm** | Role-playing, inception prompting |
| **Communication** | Turn-based dialogue |
| **Specialization** | Inception (agent creates agent) |
| **Best For** | Collaborative reasoning, research |
| **License** | Apache 2.0 |
| **Status** | Active research, many variants |

---

## SuperAGI

**Autonomous agent platform** — Full-stack with marketplace, memory, tools.

```python
from superagi.agent import SuperAgent
from superagi.tools import Toolkit

agent = SuperAgent(
    name="Researcher",
    goals=["Research AI agents", "Write report"],
    tools=Toolkit.get_default_tools()
)

await agent.run()
```

| Feature | Detail |
|---|---|
| **Paradigm** | Goal-driven, tool-rich |
| **Architecture** | Agent + Toolkit + Memory + Marketplace |
| **Memory** | Vector + Graph + Relational |
| **Best For** | Autonomous task completion |
| **License** | MIT |
| **Status** | Active, commercial (SuperAGI Cloud) |

---

## BabyAGI

**Task-driven autonomous agent** — Original viral autonomous agent.

```python
from babyagi import BabyAGI

baby = BabyAGI(
    objective="Research AI agents",
    llm=llm,
    vectorstore=vectorstore,
    max_iterations=10
)

await baby.run()
```

| Feature | Detail |
|---|---|
| **Paradigm** | Task creation → prioritization → execution |
| **Loop** | Create tasks → Prioritize → Execute → Store results |
| **Memory** | Pinecone/Chroma for context |
| **Best For** | Simple autonomous research |
| **License** | MIT |
| **Status** | Historical reference, inspired many |

---

## AutoGPT

**Original viral autonomous agent** — Recursive self-prompting.

```python
# Original concept (simplified)
while not goal_achieved:
    thought = llm.generate("What should I do next?")
    action = parse_action(thought)
    result = execute(action)
    memory.store(thought, action, result)
```

| Feature | Detail |
|---|---|
| **Paradigm** | Recursive self-prompting |
| **Memory** | Local file + vector |
| **Tools** | Web, code, file, API |
| **Best For** | Historical reference |
| **License** | MIT |
| **Status** | Superseded by modern frameworks |

---

## AgentVerse

**Multi-agent simulation** — Emergent behavior research.

```python
from agentverse import AgentVerse, Agent

agents = [
    Agent(name="Alice", role="Researcher"),
    Agent(name="Bob", role="Engineer"),
    Agent(name="Carol", role="Designer")
]

verse = AgentVerse(agents=agents)
await verse.run("Design a mobile app")
```

| Feature | Detail |
|---|---|
| **Paradigm** | Simulation, emergent behavior |
| **Communication** | Broadcast, pairwise, group |
| **Analysis** | Social network, influence, consensus |
| **Best For** | Research, social dynamics |
| **License** | MIT |
| **Status** | Research-focused |

---

## AgentBench / AgentBoard

**Benchmarking frameworks** — Standardized evaluation.

```python
from agentbench import AgentBench

bench = AgentBench()
results = bench.run(
    agent=my_agent,
    tasks=["web_shop", "data_analysis", "coding"]
)
print(results.leaderboard())
```

| Benchmark | Domains |
|---|---|
| **AgentBench** | Web, OS, DB, KG, Coding, Math |
| **SWE-bench** | Real GitHub issues (coding) |
| **GAIA** | General AI assistants (reasoning) |
| **τ-bench** | Tool use, API interaction |
| **WebShop** | E-commerce navigation |
| **ALFWorld** | Embodied, household tasks |

---

## Specialized Frameworks

### Research Agents
| Framework | Focus |
|---|---|
| **PaperQA** | Literature review, citation tracking |
| **ScholarAgent** | Hypothesis generation, experiment design |
| **ChemCrow** | Chemistry research, synthesis planning |
| **BioAgent** | Bioinformatics, genomics |

### Coding Agents
| Framework | Focus |
|---|---|
| **SWE-agent** | GitHub issue resolution |
| **OpenDevin** | End-to-end software development |
| **CodeAct** | Code as action space |
| **Aider** | Pair programming in terminal |

### Embodied / Robotics
| Framework | Focus |
|---|---|
| **SayCan** | LLM + affordance for robotics |
| **Inner Monologue** | Closed-loop robotic control |
| **RT-2** | Vision-language-action |
| **VoxPoser** | 3D manipulation from language |

### Scientific Discovery
| Framework | Focus |
|---|---|
| **AI Scientist** | End-to-end research (idea → paper) |
| **Coscientist** | Chemistry experimentation |
| **GraphReasoner** | Knowledge graph reasoning |

---

## Emerging Patterns (2024-2025)

### 1. Test-Time Compute Scaling (o1-style)
```python
# Multiple reasoning attempts, select best
class ReasoningAgent:
    async def run(self, query):
        candidates = await asyncio.gather(*[
            self.reason_once(query) for _ in range(8)
        ])
        return self.select_best(candidates)
```

### 2. Agent-Computer Interface (ACI)
```python
# Desktop control, not just API
from computer_use import ComputerTool

agent = Agent(tools=[ComputerTool()])
await agent.run("Open browser, search for flights")
```

### 3. Self-Improving / Evolutionary
```python
# Agent modifies own code/prompts
class EvolvingAgent:
    async def evolve(self, feedback):
        new_prompt = await llm.improve_prompt(self.prompt, feedback)
        self.prompt = new_prompt
        await self.save()
```

### 4. Neurosymbolic Integration
```python
# LLM + formal verification
from lean_dojo import LeanDojo

theorem = LeanDojo("theorem.lean")
proof = await llm.prove(theorem)
if theorem.verify(proof):
    return proof
```

### 5. Multi-Modal Native Agents
```python
# Native vision + audio + text
agent = Agent(
    model="gpt-4o",  # Native multimodal
    tools=[vision_tool, audio_tool, code_tool]
)
await agent.run("Watch this video and summarize")
```

---

## Tracking the Frontier

### Resources to Watch
| Source | Type |
|---|---|
| **ArXiv: cs.AI, cs.CL, cs.LG** | Daily papers |
| **Papers with Code** | Benchmarks + code |
| **Hugging Face Daily Papers** | Curated |
| **Latent Space Podcast** | Practitioner interviews |
| **AI Engineer Newsletter** | Engineering focus |
| **LangChain/Autogen/CrewAI Discords** | Community |
| **AgentBench Leaderboard** | Benchmarks |

### Conferences
| Conference | Focus |
|---|---|
| **NeurIPS / ICML / ICLR** | ML foundations |
| **ACL / EMNLP** | NLP, LLM |
| **AAAI / IJCAI** | AI, agents |
| **AgentConf** | Agent-specific |
| **ICRA / RSS** | Robotics/embodied |

---

## Evaluation Framework for Fringe Tech

| Criterion | Questions |
|---|---|
| **Maturity** | Stars, contributors, release frequency, issues |
| **Documentation** | Tutorials, API docs, examples |
| **Community** | Discord/Slack, response time, PR merge rate |
| **Production Ready** | Used in prod? By whom? Scale? |
| **Extensibility** | Plugin system, custom tools, custom models |
| **Observability** | Tracing, logging, debugging |
| **License** | MIT/Apache vs custom/restrictive |
| **Maintenance** | Last commit, open PRs, security issues |

---

## Related Notes

- [[Frameworks/Framework-Comparison.md]] — Mainstream comparison
- [[Frameworks/LangGraph.md]] — Production alternative
- [[Frameworks/AutoGen.md]] — Microsoft's multi-agent
- [[Frontiers/Research-Directions.md]] — Where these are heading
- [[Frontiers/Self-Improving-Agents.md]] — Evolutionary agents
- [[Frontiers/Agent-Computer-Interface.md]] — ACI pattern
- [[Evaluation/Agent-Benchmarks.md]] — Benchmarking
- [[Domains/Coding-Agents.md]] — SWE-agent, OpenDevin

---

*Sources: GitHub repos, ArXiv papers, framework docs, benchmark leaderboards, community Discords, conference proceedings, practitioner blogs (Lilian Weng, Andrej Karpathy, Harrison Chase, Yohei Nakajima).*