Orchestrating Autonomous Agent Teams

Resources

• How Anthropic Built Its Multi-Agent Research System — An original engineering account of lead-agent delegation, parallel researchers, context management, and evaluation. Useful for discussing which tasks benefit from isolated specialists and where coordination and token costs can outweigh parallelism. https://www.anthropic.com/engineering/multi-agent-research-system

• Claude Code Agent Teams — Documents shared task lists, dependencies, teammate messaging, separate sessions, and operator intervention. A concrete example for bidirectional coordination, progress visibility, and avoiding file collisions rather than treating a team as a batch of independent prompts. https://code.claude.com/docs/en/agent-teams

• Claude Code Subagents — Describes role definitions with separate contexts, selected tools, models, and permissions. Use it to specify bounded planner, researcher, builder, reviewer, and validator roles, and to distinguish delegated subagents from peers that coordinate with each other. https://code.claude.com/docs/en/sub-agents

• OpenAI Agents SDK Orchestration — Compares model-directed coordination with code-directed orchestration. Helps decide which routing decisions can be probabilistic and which dependency, risk, and approval boundaries should remain deterministic. https://openai.github.io/openai-agents-python/multi_agent/

• OpenAI Agents SDK Handoffs — Explains transfers of control, structured handoff inputs, and history filtering. Useful for designing explicit handoff contracts and showing why transferring to another agent is different from calling a specialist that returns a result. https://openai.github.io/openai-agents-python/handoffs/

• Git Worktrees — The official reference for multiple working trees attached to one repository, including branch restrictions and cleanup. Provides the filesystem isolation layer for concurrent builders while making clear that separate directories do not resolve semantic conflicts between changes. https://git-scm.com/docs/git-worktree

• Git Merge — Documents branch integration, conflict handling, and aborting a merge. Use it to define deliberate merge points where a reviewer or validator checks the combined result rather than accepting each agent’s branch in isolation. https://git-scm.com/docs/git-merge

• tmux Getting Started — The project’s guide to sessions, windows, panes, and attaching or detaching terminals. Useful for an operator’s view of parallel agent sessions; tiled panes improve visibility but do not replace dependency tracking, durable queues, or approval gates. https://github.com/tmux/tmux/wiki/Getting-Started

• LangGraph Workflows and Agents — Provides routing, parallelization, orchestrator–worker, and evaluator–optimizer examples. Use the explicit graph and worker-state patterns to explain fan-out, result collection, dependencies, and the point where parallel work must converge. https://docs.langchain.com/oss/python/langgraph/workflows-agents

• RouteLLM — The original routing framework provides trainable routers and evaluation tooling for choosing between stronger and cheaper models. Useful for making capability–cost tradeoffs measurable; risk permissions and latency requirements still need separate policy and measurement. https://github.com/lm-sys/RouteLLM

• Microsoft’s Agent Orchestration Patterns — Compares sequential, concurrent, handoff, group-chat, and other coordination structures with their tradeoffs. Helps select an organization based on dependencies and supervision needs, including when to retain human intervention instead of adding more agents. https://learn.microsoft.com/en-us/azure/architecture/ai-ml/guide/ai-agent-design-patterns