2.8 KiB
Agent Team Coordination
Prompt construction
Front-load all context into the initial Agent() prompt. Never send follow-up messages for information that was available at launch time. Every follow-up costs a round-trip and risks the agent having already committed to a suboptimal approach.
Bad: spawn agent, then send "also use context7 tools" Good: include "Use mcp__context7__resolve-library-id and mcp__context7__query-docs for current docs" in the original prompt
Reading selectively
When spawning agents that need to learn a pattern from existing files, point them at 2-3 structurally distinct examples — not every file. Most files in a category share 80% of their structure. Reading all N wastes context on redundancy.
Bad: "Read all 10 Python agent files to understand the pattern" Good: "Read codeflash-python.md (router), codeflash-cpu.md (domain agent), and codeflash-deep.md (orchestrator) — these three cover all structural variants"
Concise reporting
Agents report back to the team lead who has full file access. Reports should confirm completion and flag issues — not restate file contents.
Bad: 500-word summary listing every section, antipattern, and line count Good: "Done. 2 files written: codeflash-js-cpu.md (334 lines), codeflash-js-memory.md (386 lines). Both follow Python agent structure. No issues."
Only include detail when something deviated from the plan or needs the lead's attention.
Templating shared structure
When writing 3+ files with shared structure, extract the common skeleton first. Define the invariant sections (frontmatter shape, opening protocol reference, experiment loop structure, keep/discard tree, plateau detection) as a template, then fill in domain-specific content. This prevents inconsistencies in phrasing and formatting across files.
When using writer agents in parallel, include the skeleton in each agent's prompt rather than having each agent independently infer the structure from Python examples.
Tool selection hierarchy
Default to the most direct tool for the job. Spawning an agent is expensive — it allocates a full context window, loads system prompts, and returns only a summary. Use it only when the simpler options can't do the job.
Direct tool (Grep, Glob, Read, Bash + jq) ← first choice
↓ not enough
Subagent (Explore, general-purpose) ← when you need multi-step search or the target is unknown
↓ not enough
Named teammate (Agent + TeamCreate) ← when work is long-running and parallel
Concrete examples:
- Parse a known JSON file →
jqvia Bash, not an Agent - Find which files define a class → Grep, not Explore
- Search for a pattern across 200 files, then read the top 5 matches → Explore
- Profile, implement, and benchmark in parallel → named teammates