Nextop Agentic Workers

The NAW frameworkBuild, connect, run, in three steps.

1. Make workers.

   

   Anyone can make a ReAct AI worker. Set a persona, a mission, tools and work documents, and the LLM calls tools and reads the results on its own. Outputs are checked against a JSON Schema and checks.py, so only results that pass move on, and documents are searched on demand with RAG.

   • ReAct
   • Function calling
   • JSON Schema
   • checks.py
   • RAG

2. Draw the flow.

   

   Place workers as blocks and connect them to design the data flow precisely. Each line decides whose output goes to whom and which worker may open which file server. Python function blocks, loops, split / gather parallel runs and human review all fit in one flow.

   • Report / share lines
   • fs://
   • Function blocks
   • Split / gather
   • Human-in-the-loop

3. Run and look inside.

   

   Every block records its prompts, tool calls, outputs and token counts, so you find the problem in the whole flow fast. nextop check validates the structure before a run, and a failed run can resume from that block. Start small, verify, and grow it into an advanced AI service.

   • Precheck
   • Run records
   • Token usage
   • Resume

Key features

• WorkersPersona, mission, tools, outputs
• CheckersRefuse a wrong output and guide the fix
• FlowDecision, loop and parallel
• File serversShare files and knowledge (RAG)
• Start workersStart the flow and take input
• Function blocksPython steps (no LLM)
• Domain knowledgeEach worker's documents
• Human reviewStop, approve or redo
• StudioMake workers, draw flows, run and inspect
• Run recordsResume from the failed step

Principles

• The LLM writes, the program checks.A checker verifies each worker's output for format and rules, and sends it back with reasons when it is wrong. Whether it passes, and who it goes back to, is decided by code, not by the LLM.
• One worker, one job.Each worker has one narrow role and receives only the material that job needs. A shorter context costs less and keeps judgment sharp.
• Every step is traceable.Who read which material, what came out and what it cost are recorded per block, so results can be verified and reproduced.

Case study: video content automation

Shoestring Studios is a 13-worker team built with Nextop Agentic Workers. It turns a dense book chapter into a narrated explainer video.

More on YouTube ↗

Get startedFrom zero to a first run.

pip install nextop
nextop new myproject --example
nextop studio myproject
nextop run myproject --flow main

License and pricing

FAQ