Repo: https://github.com/jeongmk522-netizen/agentlas\_org\_chart
Almost every multi-agent setup I have shipped or tested eventually hits the same wall. Agents bouncing between each other, reviewers asking for one more polish pass forever, research workers spawning indefinite subtopics, tool calls spiraling until the recursion limit kicks in. The framework docs usually call these "loops" and offer a max-iteration knob. I started suspecting the knob is treating a symptom, and the real issue is closer to how the agents are organized to begin with.
The pattern that kept reappearing: when agents are designed as peers (researcher talks to analyst, analyst talks to writer, writer hands back to reviewer), nobody clearly owns the outcome. Every agent can keep asking another agent for more work. The graph has stop conditions on paper, but no single agent has the authority to declare "this is done, stop the run." That authority is implicit at best and gets diluted across the peer network.
The hypothesis I am testing is that loop failures are organization-design failures more than prompt failures. The fix is to treat the agent network as an org chart with explicit reporting lines, not a chat room of peers. One accountable mission owner. One owner per workstream. Finite delegation depth. A typed return contract per worker (status, evidence, output, blockers, next action). Manager-only authority to reopen or terminate. Memory lives at the authority layers, specialists get scoped context only.
The layers I have been working with are roughly chair, strategy office, division manager, team lead, and specialist worker, with QA and policy as separate staff offices that can reject and escalate but cannot themselves spawn unbounded new work. The reviewer-recursion failure mode in particular gets killed when verifiers are structurally allowed one reject pass, then must escalate.
Frameworks already have most of the primitives. CrewAI has a hierarchical process where a manager validates worker output. LangGraph has supervisors, subagents, and an explicit recursion limit. OpenAI Agents SDK has manager-style orchestration distinct from peer handoffs. AutoGen has GroupChatManager. Anthropic's published research system is orchestrator-worker. What I think is underused is treating the manager not as a moderator for an open group chat but as a formal reporting line with authority to terminate.
Two things I am unsure about. First, hierarchy can become its own bottleneck. If every decision routes upward, the chair agent becomes a single point of latency and a single point of failure. Second, escalation-as-feature only works if the top of the org chart has real stop authority. If the chair just calls another LLM that calls more LLMs, the loop just moved one floor up.
[link] [comments]