Orchestrators vs. Graphs: stop babysitting agents, approve only the last step
blaizedsouza · x · 2026-09-15
A widely shared X thread clarifies two agent architecture styles: an orchestrator sits above the work and routes every message — with five agents, one context reads five instructions plus five replies, bloating fast. A graph is the shape of the work itself, with no central router: a splitter hands tasks off and lets go, and a merge reads only the final results. The key line: an orchestrator sits above the work and holds all of it; a graph is the work and holds none of it. The thread links to a full course, "Loops and Graphs," covering how loops let agents self-check so humans only approve the last step — escaping babysitter mode.
More from coding & agent
- inclusionAI's LLaDA-UI: 16.7B MoE diffusion VLM for GUI agents — inclusionAI · 2026-09-15
- Andrew Chen: Agents gave me three messy inboxes to check — andrewchen · 2026-09-15
- The detective novel thought experiment: why hidden states beat tokens for model handoff — CShorten30 · 2026-09-15
- oh-my-pi open-sources agent context compaction design docs, now at 31k GitHub stars — nirmal_dist · 2026-09-15
- RAFT v3.1 open-sources retrieval-augmented fine-tuning to clone human personas — jessi_cata · 2026-09-15
- 18 sub-agents cut to 1: Polylane argues sub-agents are just wrong — zeeg · 2026-09-15