Digital Product Engineering6.6 Agent Orchestration & MCP
VOL. VI · CH. 6.6 · AI SYSTEMS

Agent Orchestration & MCP

What happens when a task needs more than one tool, one model, or one system to complete.

DivisionAI Engineering
DifficultyAdvanced
Prerequisites6.5
Related6.7 1.5
2 min read · 360 words

6.6.1Definition

Orchestration coordinates multiple tools, sub-agents, or steps toward a single larger goal, managing sequencing, error handling, and hand-offs between them. MCP (Model Context Protocol) is an open standard that lets a model connect to external tools and data sources through a common interface, rather than each integration being built as a one-off.

6.6.2Why It Exists

As agentic systems (6.5) grow beyond a single tool call, coordinating multiple tools, sub-tasks, and error paths becomes its own engineering problem distinct from any individual tool call. MCP specifically exists to standardize how models discover and call external tools, so a new integration (a calendar, a database, a third-party API) can be added once and reused across any MCP-compatible agent, rather than rebuilding the connection logic for every model or application separately.

6.6.3Orchestration Patterns

PatternStructureBest fit
Single agent, many toolsOne model with a broad tool set (6.5)Most product features — simplest to reason about
Sequential sub-agentsOutput of one agent step feeds the nextMulti-stage pipelines (research → draft → review)
Supervisor / workerA coordinating agent delegates to specialized sub-agentsComplex tasks spanning distinct domains

6.6.4Common Mistakes

  • Over-engineering orchestration for a task a single agent with a few tools would handle, adding coordination complexity with no corresponding capability gain.
  • No shared error-handling strategy across orchestrated steps, so a failure in one sub-agent silently corrupts or stalls the overall task.
  • Building a custom, one-off tool-connection protocol instead of adopting an existing standard like MCP, duplicating integration work for every new tool and every new model.

6.6.5Best Practices

  • Default to a single agent with well-defined tools; escalate to multi-agent orchestration only when a task genuinely spans distinct domains or requires specialization.
  • Adopt MCP or an equivalent standard for tool connections rather than custom integration code per tool.
  • Define explicit error and retry behavior at every hand-off point in an orchestrated pipeline.
Real-World ExampleClaude's MCP connectors — to Google Drive, Slack, Supabase, and dozens of others — let one agent access many external systems through a single standardized protocol, rather than Anthropic or every developer building a bespoke integration for each service individually.