explainer
Agent vs workflow
Three public systems, three autonomy levels. Not everything needs to be an agent.
The mix-up
“Agent” is too often a marketing label for a pipeline with an LLM in the middle. This Lab keeps workflow, tool use and evaluable agent apart on purpose.
Three pieces of evidence, three shapes
- Knowledge Copilot — retrieval then generation. A grounded workflow, not an action loop in the world.
- Automation Agent — classification, extraction, score, human in the loop. Semi-deterministic. The README is explicit: not a multi-agent system.
- Decision Copilot — profiling → investigation → scenarios. More agentic. The profiling lab has a versioned A/B/C table (sample vs tool vs claims).
Rule
If failure is invisible (no tool, no metric, no review), it is copy, not a system. An honest workflow beats a cosmetic agent.
What I am not concluding
No “best framework” ranking. No MCP claim. The tabular profiling A/B/C table and the HITL gold (hitl-agent.v2) are published.