An agent is an autonomous unit that uses a language model, tools, and instructions to perform tasks and make decisions.
2
Multi-agent systems divide complex workflows among specialized agents that can work in parallel, in series, or under a hierarchy.
3
A generalist agent framework could create and coordinate agents from templates, but it raises hard questions about evaluation, monitoring, fine-tuning, privacy, and orchestration.
Summary
Natan describes agents as autonomous units that use language models, tools, and system instructions to perform tasks. Teams of agents can divide work such as outreach, grant applications, financial research, or event planning into tasks that run in parallel, in sequence, or under hierarchical control. He became interested in this approach after building a chatbot that struggled with large collections of files and complex, domain-specific questions. His proposed open-source project focuses on a general-purpose framework with reusable agent templates, domain-specific agents, tool access, orchestration, logging, and monitoring. An orchestrator would interpret a user request, choose the necessary agents, arrange their order, assign tools, and combine their results. Natan is clear that this is difficult in practice. The open questions include how to evaluate agent systems, improve the orchestrator, fine-tune individual modules, run agents privately, and monitor their behavior. The project is intended as a shared place for developers to experiment with these systems.
Agents combine language models, tools, and instructions to perform autonomous tasks
Natan defines an agent as an autonomous unit that can perform tasks and make decisions. A system prompt describes what it should do, while tools let it search the web, use a computer, or write and run code. The underlying language model might be GPT, Claude, Llama, or Mistral. A crew is a team of agents with different responsibilities, similar to product managers, software engineers, writers, researchers, and salespeople working together.
Multi-agent workflows can run in parallel, in sequence, or through hierarchy
A crew can divide work in several ways. Tasks can run in parallel when they do not depend on one another. They can also run in series, with one result passed to the next task. A hierarchical workflow can use conditional logic, such as deciding what to do after an earlier step. Natan uses the word workflow for the complete end-to-end process.
Agentic retrieval emerged from the limits of a file-question-answering chatbot
While building a chatbot for uploaded files at anote, Natan found that large collections and complex, domain-specific questions were difficult to answer well. The team considered fine-tuning, enhanced retrieval-augmented generation, labeling, and different evaluations. They then explored agentic RAG, where a flow of agents can break down the work and use different tools or steps to produce an answer.
Agents can automate workflows such as outreach, grant applications, finance research, and events
For outreach, agents could find people, enrich information from the web, generate emails, and send them. A grant workflow could research opportunities, write a tailored application, check its format, revise it, and submit it through a portal. A finance workflow could combine a web-search agent with a finance agent that gathers stock information. Event planning could assign venue search, logistics, and marketing to different agents.
A generalist framework would create agents from reusable templates
Natan contrasts specific agents built for a single use case with a generalist framework. In the generalist design, templates provide agents with access to tools. A user submits a request, and an orchestrator decides which templates to use and how to arrange them. The resulting agents can work in series or in parallel, complete their assigned tasks, and return values to the wider system.
The orchestrator is the central control point for autonomous multi-agent work
The orchestrator would interpret a request such as contacting AI leaders about an event. It would spawn the needed agents, assign their order, provide their tools and tasks, and coordinate their results. In the later explanation, Natan says agents could be created or edited through YAML or Markdown, with their work monitored and logged. The intended interface could be a chatbot or developer-facing code.
Generalist agents remain difficult to evaluate and operate reliably
Natan describes the generalist approach as attractive in theory but difficult in practice. The open questions include how to optimize the orchestrator, evaluate agentic systems for reliability, fine-tune individual modules, decide what training data is needed, run agents privately, and monitor them through logs. These issues are part of the proposed open-source research project.
The open-source project aims to give developers a place to experiment with agent teams
Natan plans to open-source the project so people can learn how agents work together. The proposed framework includes a general-purpose agent system, a registry, domain-specific agents, smart orchestration, tool use, logging, and monitoring. A standalone chatbot and setup instructions were already available for people to clone and try. The project is intended to help the team and the wider community learn through implementation.
"This orchestrator is going to basically be able to take in this query, figure out which agents or templates it should call, and what order they should actually do the operations."13:21
Who should watch
You are building workflows that involve research, writing, web access, or several dependent steps and want to understand where multiple agents might fit.
You are deciding between specialized agents and a general-purpose agent framework, especially around orchestration and reusable templates.
You need to think through evaluation, logging, privacy, and fine-tuning before putting autonomous agents into a real system.
