Skip to content

Agents

Agents are the orchestration layer. Where a skill does one job well, an agent strings several skills together and picks the right model for the work. You rarely call an agent by name: commands and workflows summon them for you. This page is the map of who does what.

AgentRoleKey skillsModel
Research AgentLiterature search and synthesisliterature-search, citation-managementinherit
Writing AgentSection drafting and coherencepaper-drafting, writing-style-analysis, figure-suggestionsinherit
Review AgentMulti-perspective peer reviewpeer-reviewinherit
Formatting AgentLaTeX and Word formatting and compliancejournal-formatting, latex-tables, tikz-diagrams, word-outputsonnet
Code AgentCode analysis and implementationimplementation, code-analysissonnet
Style AgentVoice calibrationwriting-style-analysisinherit
Visualization AgentPlots, diagrams, and NN architecturesvisualization, tikz-diagrams, plotneuralnet, figure-suggestions, image-prompt-craftingsonnet
Statistics AgentMethod selection and experiment designstatistical-analysis, experiment-design, visualizationinherit
Discovery AgentJournal and conference finding, gaps, and SOTAjournal-finder, conference-finder, research-gaps, sota-finderinherit

The Code, Visualization, and Formatting agents run on Sonnet. Their work is mechanical and high-volume (generating figures, compiling LaTeX, producing tables), so Sonnet handles it faster and cheaper without a loss in quality. That keeps the larger model’s budget for the Research and Writing agents, where synthesis, argument, and voice actually benefit from the extra reasoning. An inherit model simply means the agent uses whatever model the parent session is running, so your top-level choice flows down to the reasoning-heavy work.

Agent definitions carry two independent frontmatter mechanisms. A model: sonnet key pins the agent to Sonnet, as described above. A separate skills: key preloads the listed skills into that agent’s context so they are available without a discovery step. These do not make skills route into an agent on their own: a skill only forks into an agent when the skill’s own frontmatter declares context: fork together with agent: <name>. Today that happens for two skills: implementation and code-analysis, both of which fork into the Code Agent.

  • The skills each agent orchestrates are catalogued in Skills.
  • Commands that summon these agents live in Commands.
  • To see the Research and Writing agents work end to end, follow the guides.