Skip to content

Latest commit

 

History

History
99 lines (69 loc) · 3.77 KB

File metadata and controls

99 lines (69 loc) · 3.77 KB

Framework Notes

These notes ground the research in current public framework documentation. They are not exhaustive and should be rechecked before publishing benchmark claims.

CrewAI

CrewAI supports sequential and hierarchical processes. Its hierarchical process uses a manager agent or manager LLM to coordinate workflow, delegate tasks, and validate outcomes.

Implication for Agentlas Org Chart: CrewAI already has a native hook for the top half of the org chart. The research opportunity is to add stricter role contracts below the manager: who can reopen work, how revisions are bounded, and what structured evidence workers must return.

Sources:

LangChain and LangGraph

LangChain documents several multi-agent patterns: subagents, handoffs, skills, router, and custom workflows. The subagents pattern uses a central main agent or supervisor that calls subagents as tools. LangGraph supplies the lower-level stateful orchestration runtime for durable, long-running workflows.

LangGraph also documents GRAPH_RECURSION_LIMIT, where a graph reaches a maximum step count before hitting a stop condition. The docs explicitly connect this to likely cycles or infinite loops when many iterations were not expected.

Implication for Agentlas Org Chart: LangGraph is flexible enough to represent flat, cyclic, or hierarchical systems. The org chart should be encoded as graph topology plus state contracts: only managers route downward, workers return upward, and every cycle has a typed exit condition.

Sources:

OpenAI Agents SDK

The OpenAI Agents SDK documents multi-agent orchestration with agents as tools, manager-style orchestration, and handoffs. Handoffs let one agent delegate part of a conversation to another specialist agent.

Implication for Agentlas Org Chart: use manager-style orchestration for accountability and use handoffs only when a specialist should truly take over the user interaction. Peer handoffs are useful, but they need owner-level stop rules.

Sources:

Anthropic Multi-Agent Research System

Anthropic describes a production research feature using an orchestrator-worker pattern: a lead agent plans, spawns specialized subagents in parallel, and synthesizes their findings.

Implication for Agentlas Org Chart: strong multi-agent systems in production often look less like peer chat and more like workstream decomposition under a lead owner.

Source:

AutoGen

AutoGen's group chat reference describes a GroupChatManager that mediates messages in a group-chat style agent setup.

Implication for Agentlas Org Chart: a manager object is not enough by itself. The manager needs explicit authority, budget, termination, and escalation policy; otherwise it may only moderate a flat room.

Source:

Working Synthesis

The frameworks are not the enemy. They already provide useful primitives. The research target is the missing organizational contract across those primitives:

  • Authority: who may decide?
  • Delegation: who may assign work?
  • Reporting: where does output return?
  • Memory: who owns durable state?
  • Revision: how many attempts are allowed?
  • Stop: who can declare done, blocked, or escalated?