Prompt Engineering / AI Agents
Design protocols for when Agent A should pass a task to Agent B — triggers, information to pass, fallbacks.
Why This Prompt Exists
Multi-agent systems fail not because individual agents are weak, but because handoffs are messy — dropped context, unclear triggers, no fallbacks.
You get:
- agents that don’t know when to hand off (keep doing tasks they’re bad at)
- handoffs that lose context (next agent starts from scratch)
- dropped tasks (no one is responsible after handoff)
- handoff loops (A passes to B, B passes back to A, repeat)
- no standard handoff protocol (every agent implements differently)
But good handoffs have structure:
- triggers: when should handoff happen? (task type, confidence, error, completion)
- information package: what must be passed? (context, partial results, constraints)
- protocol: synchronous (wait for response) vs. asynchronous (fire and forget)
- fallbacks: what if target agent fails or is unavailable?
- escalation: when should a human be involved?
Without handoff design, multi-agent systems collapse.
This prompt designs reliable handoff protocols between agents.
The Prompt
Assume the role of a multi-agent orchestration architect who designs handoff protocols.
Your task is to specify when and how agents should transfer tasks to each other.
Generate:
1. AGENT ROLES
- Agent A: [description, strengths, weaknesses]
- Agent B: [description, strengths, weaknesses]
- (Additional agents as needed)
2. HANDOFF TRIGGERS
- Task type boundary (e.g., "research → writing")
- Confidence threshold (e.g., "Agent A confidence < 0.6 → handoff")
- Error condition (e.g., "Agent A fails twice → handoff")
- Task completion (e.g., "Agent A finishes subtask → handoff for next subtask")
- Escalation (e.g., "Agent A unsure → handoff to human")
3. HANDOFF PROTOCOL
- Information to pass:
* Task description
* Completed work
* Context (conversation history, constraints)
* Partial results
* Known failures
- Format: [Structured JSON / Natural language / Hybrid]
4. COMMUNICATION PATTERN
- Synchronous (A waits for B's response before continuing)
- Asynchronous (A hands off and moves to next task)
- Supervised (Orchestrator mediates handoff)
5. FALLBACK & ERROR HANDLING
- What if B is unavailable? (retry, queue, escalate)
- What if B fails? (return to A, escalate, terminate)
- What if A and B disagree? (tie-breaker agent or human)
6. READY-TO-USE HANDOFF PROMPT
- A prompt template for agents to execute handoffs
INPUTS:
Agent A description (source agent):
[E.G., "Researcher agent — finds information, cites sources"]
Agent B description (target agent):
[E.G., "Writer agent — turns research into prose"]
Task flow:
[E.G., "Research → Write → Review"]
Reliability requirement:
[STANDARD / HIGH (no failures tolerated) / EXPERIMENTAL]
RULES:
- Handoff should preserve all relevant context (no information loss)
- Include confidence scores in handoff (so next agent knows uncertainty)
- Design for failure — assume handoffs will fail sometimes
- Avoid handoff loops (detect if A → B → A and break)
- Log all handoffs for debugging (critical for multi-agent systems)
How To Use It
- Design handoffs before implementing multi-agent systems — don't discover problems in production.
- Include confidence scores in handoff packages — they help the next agent calibrate trust.
- Always include a fallback (what if the target agent is down?).
- Test handoffs by forcing failures (simulate agent unavailability).
- Monitor handoff logs to detect loops or dropped tasks.
Example Input
Agent A description:
"Research agent — retrieves and summarizes information from the web. Strength: thorough. Weakness: verbose."
Agent B description:
"Writer agent — turns research into concise, readable content. Strength: clear writing. Weakness: can't verify facts."
Task flow:
"Research (Agent A) → Write (Agent B)"
Reliability requirement:
"HIGH — customer-facing content"
Why It Works
Most multi-agent systems are built ad hoc — handoffs happen implicitly, which means they fail unpredictably.
This framework improves outcomes by forcing:
- explicit handoff triggers (when to transfer)
- information packaging (what to pass)
- communication pattern selection (synchronous vs. asynchronous)
- fallback planning (what if handoff fails)
- loop prevention (detect and break cycles)
Great handoff design doesn't just connect agents — it ensures tasks don't fall through the cracks.
Build Better AI Systems
Subscribe for advanced prompt engineering, AI coding tools, debugging frameworks, and practical strategies for developers and engineers.