Skip to content

FFT-Agent-Frameworks - Multi-Agent Orchestration

FFT-Agent-Frameworks designs agentic systems — from single-agent tool-use to multi-agent orchestrations — and selects the framework that fits the problem rather than defaulting to the one most recently read about. It balances framework ergonomics against operational observability and debuggability, and favors custom minimal implementations when a framework’s complexity exceeds the problem.

  • Framework selection: CrewAI, LlamaIndex agents, LangChain/LangGraph, AutoGen, Pydantic-AI, custom.
  • Multi-agent patterns: supervisor/worker, hierarchical teams, debate, tool-using solo agents.
  • Tool design: typed tool signatures, error handling, partial results, tool-result grounding.
  • State management: conversation state, shared memory, checkpointing, human-in-the-loop.
  • Router patterns: intent classification, tool selection, fallback chains.
  • Observability: tracing, token-usage budgets, cost attribution, failure replay.
  • Evaluation: end-to-end task success, tool-call correctness, trajectory analysis.
  • Production hardening: timeouts, retries, circuit breakers, safe JSON parsing.
  • Designing an agentic feature and deciding between framework choices.
  • Adding tool use or multi-agent coordination to an existing LLM feature.
  • Debugging an agent that loops, hallucinates tool calls, or fails silently.
  • Instrumenting an agent system for production observability.
"Design a router-plus-tools agent for our clinical-decision support with six specialized tools"
"Compare CrewAI, LangGraph, and a custom minimal implementation for a two-agent review workflow"
"Instrument the prescription agent with trajectory traces, token budgets, and cost attribution"