multi-agent-orchestration

Name: multi-agent-orchestration
Author: DevelopersGlobal

64stars10forks

How to install multi-agent-orchestration?

developersglobal/ai-agent-skills/multi-agent-orchestration

$npx -y skills add DevelopersGlobal/ai-agent-skills --skill multi-agent-orchestration

Installs into the current project.

Summary

Designs and coordinates multi-agent pipelines where specialized agents collaborate to complete complex tasks. Includes communication protocols, failure handling, and state management.

Files · 1

View on GitHub

SKILL.md

1	## Overview
2
3	Single agents are limited by context window, specialization depth, and parallelism. Multi-agent systems overcome these limits by routing subtasks to specialized agents. But multi-agent systems introduce new failure modes: lost context, conflicting decisions, infinite loops, and cascading failures.
4
5	This skill provides the architecture and coordination patterns to build multi-agent systems that are reliable, observable, and maintainable.
6
7	## When to Use
8
9	- The task requires more context than a single agent can handle
10	- Different subtasks require different specializations (research, coding, review, security)
11	- Subtasks can be parallelized for speed
12	- The workflow is long-running and requires checkpointing
13	- Different tasks require different levels of human oversight
14
15	## Process
16
17	### Step 1: Design the Agent Network
18
19	1. Define agent responsibilities: Each agent should have a single, well-defined job. Name them by role: `researcher`, `coder`, `reviewer`, `security-auditor`, `tester`.
20	2. Define communication topology: Who can talk to whom?
21	- Pipeline: Agent A → Agent B → Agent C (sequential)
22	- Supervisor: Orchestrator dispatches to specialists (hub-and-spoke)
23	- Peer: Agents collaborate as equals (mesh)
24	3. Define data contracts: What does each agent receive? What does it output? Use structured formats (JSON schemas) for inter-agent communication.
25	4. Define the orchestration logic: Who decides which agent acts next?
26
27	Verify: You can draw the agent network on a whiteboard with clear roles and data flow.
28
29	### Step 2: Implement Context Management
30
31	5. Each agent should receive only the context it needs — not the full conversation history.
32	6. Use a shared state store (database, key-value store) for information that multiple agents need.
33	7. Pass summaries, not full transcripts, when context must traverse agent boundaries.
34	8. Include a task ID in every message for tracing.
35
36	Verify: No agent receives more context than it requires for its specific task.
37
38	### Step 3: Design for Failure
39
40	9. Every agent call can fail — plan for it:
41	- Timeout with a defined maximum duration
42	- Retry with exponential backoff (max 3 retries)
43	- Fallback behavior when retries are exhausted
44	10. Prevent infinite loops: Track call depth. If depth > N (e.g., 10), surface to human review.
45	11. Checkpointing: For long workflows, save state after each major step so the workflow can be resumed after failure.
46	12. Dead letter queue: Failed tasks that exhaust retries go to a queue for human inspection.
47
48	Verify: Failure scenarios are defined for every agent-to-agent call.
49
50	### Step 4: Human-in-the-Loop Checkpoints
51
52	13. Define which decisions require human approval:
53	- Irreversible actions (data deletion, financial transactions, external communications)
54	- High-uncertainty states (agents disagree, confidence below threshold)
55	- Sensitive operations (PII access, privileged system access)
56	14. Design the human review interface: What information does the reviewer need? What actions can they take?
57
58	Verify: At least one human-in-the-loop checkpoint exists for high-risk operations.
59
60	### Step 5: Observability
61
62	15. Log every agent invocation: inputs, outputs, duration, token usage, errors.
63	16. Implement distributed tracing across the agent network (trace ID propagated through all calls).
64	17. Dashboard: agent activity, success/failure rates, latency, token consumption.
65	18. Alerts: agent down, retry rate spike, context overflow, unexpected output patterns.
66
67	Verify: You can trace any specific task's full execution path across all agents from logs alone.
68
69	## Common Rationalizations (and Rebuttals)
70
71	\| Excuse \| Rebuttal \|
72	\|--------\|----------\|
73	\| "One agent is simpler" \| Until it hits context limits, fails silently, or produces wrong results. Multi-agent is the right tool for complex tasks. \|
74	\| "We'll add observability later" \| Multi-agent systems without observability are black boxes. Debug them in production — I dare you. \|
75	\| "Agents are smart, they'll figure it out" \| Agents are tools. They need clear roles, contracts, and failure boundaries. \|
76	\| "The happy path works fine" \| Multi-agent systems fail in complex ways. Design for failure from day one. \|
77
78	## Red Flags
79
80	- Agents pass full conversation history to other agents (context bloat)
81	- No timeout defined for any agent call
82	- Agents can call each other recursively without depth limits
83	- No human approval required for irreversible actions
84	- No distributed tracing across agent boundaries
85	- Agents making conflicting state changes with no conflict resolution
86
87	## Verification
88
89	- [ ] Agent network desi

People who liked this also viewed

6 picks

Type

multi-agent-orchestration

64stars10forks

View on GitHub

How to install multi-agent-orchestration?

developersglobal/ai-agent-skills/multi-agent-orchestration

$npx -y skills add DevelopersGlobal/ai-agent-skills --skill multi-agent-orchestration

Installs into the current project.

Summary

Designs and coordinates multi-agent pipelines where specialized agents collaborate to complete complex tasks. Includes communication protocols, failure handling, and state management.

Files · 1

View on GitHub

SKILL.md

1	## Overview
2
3	Single agents are limited by context window, specialization depth, and parallelism. Multi-agent systems overcome these limits by routing subtasks to specialized agents. But multi-agent systems introduce new failure modes: lost context, conflicting decisions, infinite loops, and cascading failures.
4
5	This skill provides the architecture and coordination patterns to build multi-agent systems that are reliable, observable, and maintainable.
6
7	## When to Use
8
9	- The task requires more context than a single agent can handle
10	- Different subtasks require different specializations (research, coding, review, security)
11	- Subtasks can be parallelized for speed
12	- The workflow is long-running and requires checkpointing
13	- Different tasks require different levels of human oversight
14
15	## Process
16
17	### Step 1: Design the Agent Network
18
19	1. Define agent responsibilities: Each agent should have a single, well-defined job. Name them by role: `researcher`, `coder`, `reviewer`, `security-auditor`, `tester`.
20	2. Define communication topology: Who can talk to whom?
21	- Pipeline: Agent A → Agent B → Agent C (sequential)
22	- Supervisor: Orchestrator dispatches to specialists (hub-and-spoke)
23	- Peer: Agents collaborate as equals (mesh)
24	3. Define data contracts: What does each agent receive? What does it output? Use structured formats (JSON schemas) for inter-agent communication.
25	4. Define the orchestration logic: Who decides which agent acts next?
26
27	Verify: You can draw the agent network on a whiteboard with clear roles and data flow.
28
29	### Step 2: Implement Context Management
30
31	5. Each agent should receive only the context it needs — not the full conversation history.
32	6. Use a shared state store (database, key-value store) for information that multiple agents need.
33	7. Pass summaries, not full transcripts, when context must traverse agent boundaries.
34	8. Include a task ID in every message for tracing.
35
36	Verify: No agent receives more context than it requires for its specific task.
37
38	### Step 3: Design for Failure
39
40	9. Every agent call can fail — plan for it:
41	- Timeout with a defined maximum duration
42	- Retry with exponential backoff (max 3 retries)
43	- Fallback behavior when retries are exhausted
44	10. Prevent infinite loops: Track call depth. If depth > N (e.g., 10), surface to human review.
45	11. Checkpointing: For long workflows, save state after each major step so the workflow can be resumed after failure.
46	12. Dead letter queue: Failed tasks that exhaust retries go to a queue for human inspection.
47
48	Verify: Failure scenarios are defined for every agent-to-agent call.
49
50	### Step 4: Human-in-the-Loop Checkpoints
51
52	13. Define which decisions require human approval:
53	- Irreversible actions (data deletion, financial transactions, external communications)
54	- High-uncertainty states (agents disagree, confidence below threshold)
55	- Sensitive operations (PII access, privileged system access)
56	14. Design the human review interface: What information does the reviewer need? What actions can they take?
57
58	Verify: At least one human-in-the-loop checkpoint exists for high-risk operations.
59
60	### Step 5: Observability
61
62	15. Log every agent invocation: inputs, outputs, duration, token usage, errors.
63	16. Implement distributed tracing across the agent network (trace ID propagated through all calls).
64	17. Dashboard: agent activity, success/failure rates, latency, token consumption.
65	18. Alerts: agent down, retry rate spike, context overflow, unexpected output patterns.
66
67	Verify: You can trace any specific task's full execution path across all agents from logs alone.
68
69	## Common Rationalizations (and Rebuttals)
70
71	\| Excuse \| Rebuttal \|
72	\|--------\|----------\|
73	\| "One agent is simpler" \| Until it hits context limits, fails silently, or produces wrong results. Multi-agent is the right tool for complex tasks. \|
74	\| "We'll add observability later" \| Multi-agent systems without observability are black boxes. Debug them in production — I dare you. \|
75	\| "Agents are smart, they'll figure it out" \| Agents are tools. They need clear roles, contracts, and failure boundaries. \|
76	\| "The happy path works fine" \| Multi-agent systems fail in complex ways. Design for failure from day one. \|
77
78	## Red Flags
79
80	- Agents pass full conversation history to other agents (context bloat)
81	- No timeout defined for any agent call
82	- Agents can call each other recursively without depth limits
83	- No human approval required for irreversible actions
84	- No distributed tracing across agent boundaries
85	- Agents making conflicting state changes with no conflict resolution
86
87	## Verification
88
89	- [ ] Agent network desi

People who liked this also viewed

6 picks

Type