Anthropic's Workflow-First Approach to Agent Tool Design
TRIGGER
AI agents were consuming excessive context by making sequential tool calls for common workflows—each intermediate result took up limited context space, and agents frequently lost track of multi-step processes or made errors chaining calls together.
APPROACH
Anthropic's team consolidated frequently-chained operations into single tools that handle multiple discrete API calls internally. Instead of implementing `list_users`, `list_events`, and `create_event` separately, they built a `schedule_event` tool that finds availability and schedules in one call. Instead of `get_customer_by_id`, `list_transactions`, and `list_notes`, they built `get_customer_context` that compiles all relevant customer information at once. Instead of `read_logs`, they built `search_logs` that returns only relevant lines with surrounding context.
PATTERN
“Agents lose track chaining calls and burn context on intermediate results. Design tools around workflows agents actually pursue, not your REST endpoints. Five tools returning 500 tokens each when one returns 600.”
✓ WORKS WHEN
- Tool call sequences follow predictable patterns that can be identified from evaluation transcripts
- Intermediate outputs from chained calls don't need agent judgment before proceeding
- Agents frequently make errors or get confused chaining specific tool sequences together
- Context consumption from intermediate results exceeds 1000+ tokens per workflow
- The consolidated workflow maps to a recognizable human task (schedule meeting, get customer context)
✗ FAILS WHEN
- Workflows require agent judgment between steps (review results before proceeding)
- The operations being consolidated serve genuinely different use cases that shouldn't be coupled
- Users need visibility into intermediate steps for debugging or audit purposes
- The consolidated tool would require >10 parameters to handle all workflow variations
- Tool call patterns vary significantly across different task types in your evaluation