agent-fungibility

Name: agent-fungibility
Author: Dicklesworthstone

51installs13forks

How to install agent-fungibility?

dicklesworthstone/agent_flywheel_clawdbot_skills_and_integrations/agent-fungibility

$npx -y skills add Dicklesworthstone/agent_flywheel_clawdbot_skills_and_integrations --skill agent-fungibility

Installs into the current project.

Summary

The philosophy and practical benefits of agent fungibility in multi-agent software development. Why homogeneous, interchangeable agents outperform specialized role-based systems at scale.

Files · 1

View on GitHub

SKILL.md

1	# Agent Fungibility — The Key to Scalable Agent Swarms
2
3	> Core Principle: "Fungibility bestows a LOT of really good properties automatically in a computer system."
4	>
5	> YOU are the bottleneck. Be the clockwork deity to your agent swarms: design a beautiful and intricate machine, set it running, and then move on to the next project.
6
7	---
8
9	## The Debate: Specialized vs. Fungible Agents
10
11	A burgeoning debate has sprung up around the optimal way to scale up the number of agents working on a software project:
12
13	\| Approach \| Description \|
14	\|----------\|-------------\|
15	\| Specialized Roles \| Assign distinct roles (tester agent, backend agent, committer agent, etc.) \|
16	\| Fungible Agents \| All agents are identical and can do any task \|
17
18	The answer for software development: Fungible agents win.
19
20	---
21
22	## Why Fungibility Works
23
24	### The Problem with Specialized Agents
25
26	Rather than setting up a complex heterogeneous system where you have assigned roles to specific tasks for agents, which is brittle and creates problems when one of the agents crashes or loses its memory or otherwise suffers from context rot and needs to be euthanized, you have to deal with:
27
28	1. Understanding what kind of agent just died
29	2. What was it doing when it died
30	3. How to replace it functionally with the relevant context
31	4. What happens to the other agents that were depending on that specialized agent?
32
33	### The Fungibility Solution
34
35	You're better off just having a bunch of fungible agents that are all simply executing beads of any kind.
36
37	They can be any kind of agent:
38	- Claude Code
39	- Codex
40	- Antigravity CLI (`agy`, pinned to "Gemini 3.1 Pro (High)"; successor to the retiring Gemini CLI)
41	- Amp
42	- Cursor
43	- etc.
44
45	But they're fungible in that they can all wear any hat and adaptively assume any role depending on the bead task they're working on.
46
47	---
48
49	## The Two Big Unlocks
50
51	### 1. Use BV to Choose Optimal Next Bead
52
53	Each agent uses `bv --robot-triage` or `bv --robot-next` to find the highest-impact ready bead. No central assignment needed.
54
55	### 2. Put Everything in the Plan → Beads
56
57	If you put everything you want to have happening in your original markdown plan that you turn into beads, you get everything else "for free" and it's robust to bad things happening to any agents.
58
59	---
60
61	## Robustness Properties
62
63	Agents fail. They:
64	- Crash
65	- You accidentally close the tab
66	- Context rot (lose track of what they're doing)
67	- Get stuck in loops
68	- Need to be "euthanized" and restarted
69
70	With fungible agents, you never have to worry about that stuff.
71
72	When an agent dies:
73	1. The bead it was working on remains marked in-progress
74	2. Any other agent can pick it up
75	3. No special replacement logic needed
76	4. No dependency on that specific agent
77
78	---
79
80	## Scaling Properties
81
82	When you want to move faster:
83	- Simply spin up more agents at any time
84	- Don't need to think about whether there's an imbalance between the counts of each agent role
85	- 3 agents → 10 agents → 20 agents: just add more, they all do the same thing
86
87	Agent fungibility lets you go much faster and scales better, especially as:
88	- Projects get larger and more complex
89	- Agent counts increase past 10 working on the same project at the same time
90
91	---
92
93	## The Analogy: Fountain Codes
94
95	This is why I'm a huge fan of things like fountain codes (like RaptorQ) for file storage:
96
97	> Turn a file into an endless stream of fungible blobs with very little overhead and the user can catch any blob in any order and each new blob helps them reconstruct the file; there's no "rarest chunk" like with BitTorrent.
98
99	In software development:
100	- Each bead is a "blob"
101	- Any agent can work on any bead
102	- There's no "critical specialist" that becomes a bottleneck
103	- The system is resilient to partial failures
104
105	---
106
107	## When Specialized Roles DO Make Sense
108
109	There are contexts where designated roles for agents work better. Example: Automated scientific inquiry (like BrennerBot).
110
111	Why it works there:
112	- The discourse itself (the back and forth between different agent types) is the core mechanism of surfacing truth
113	- The debate structure IS the point
114
115	Why it doesn't work for software:
116	- In software development, we just want good code that works
117	- The goal is output, not discourse
118	- Fungibility removes bottlenecks and single points of failure
119
120	---
121
122	## The Clockwork Deity Philosophy
123
124	> Most of my agent tooling and workflows are about removing me from the equation, because there's only one of me, but a potentially unlimited number of agents.
125
126	The way to do this:
127	1. Front-load all input in planning phases — Use GPT Pro with Extended Reasoning, multiple iterations, blend feedback from all frontier models
128	2. Offload task structuring — Convert plans to beads with full d

People who liked this also viewed

6 picks

Type

agent-fungibility

51installs13forks

View on GitHub

How to install agent-fungibility?

dicklesworthstone/agent_flywheel_clawdbot_skills_and_integrations/agent-fungibility

$npx -y skills add Dicklesworthstone/agent_flywheel_clawdbot_skills_and_integrations --skill agent-fungibility

Installs into the current project.

Summary

The philosophy and practical benefits of agent fungibility in multi-agent software development. Why homogeneous, interchangeable agents outperform specialized role-based systems at scale.

Files · 1

View on GitHub

SKILL.md

1	# Agent Fungibility — The Key to Scalable Agent Swarms
2
3	> Core Principle: "Fungibility bestows a LOT of really good properties automatically in a computer system."
4	>
5	> YOU are the bottleneck. Be the clockwork deity to your agent swarms: design a beautiful and intricate machine, set it running, and then move on to the next project.
6
7	---
8
9	## The Debate: Specialized vs. Fungible Agents
10
11	A burgeoning debate has sprung up around the optimal way to scale up the number of agents working on a software project:
12
13	\| Approach \| Description \|
14	\|----------\|-------------\|
15	\| Specialized Roles \| Assign distinct roles (tester agent, backend agent, committer agent, etc.) \|
16	\| Fungible Agents \| All agents are identical and can do any task \|
17
18	The answer for software development: Fungible agents win.
19
20	---
21
22	## Why Fungibility Works
23
24	### The Problem with Specialized Agents
25
26	Rather than setting up a complex heterogeneous system where you have assigned roles to specific tasks for agents, which is brittle and creates problems when one of the agents crashes or loses its memory or otherwise suffers from context rot and needs to be euthanized, you have to deal with:
27
28	1. Understanding what kind of agent just died
29	2. What was it doing when it died
30	3. How to replace it functionally with the relevant context
31	4. What happens to the other agents that were depending on that specialized agent?
32
33	### The Fungibility Solution
34
35	You're better off just having a bunch of fungible agents that are all simply executing beads of any kind.
36
37	They can be any kind of agent:
38	- Claude Code
39	- Codex
40	- Antigravity CLI (`agy`, pinned to "Gemini 3.1 Pro (High)"; successor to the retiring Gemini CLI)
41	- Amp
42	- Cursor
43	- etc.
44
45	But they're fungible in that they can all wear any hat and adaptively assume any role depending on the bead task they're working on.
46
47	---
48
49	## The Two Big Unlocks
50
51	### 1. Use BV to Choose Optimal Next Bead
52
53	Each agent uses `bv --robot-triage` or `bv --robot-next` to find the highest-impact ready bead. No central assignment needed.
54
55	### 2. Put Everything in the Plan → Beads
56
57	If you put everything you want to have happening in your original markdown plan that you turn into beads, you get everything else "for free" and it's robust to bad things happening to any agents.
58
59	---
60
61	## Robustness Properties
62
63	Agents fail. They:
64	- Crash
65	- You accidentally close the tab
66	- Context rot (lose track of what they're doing)
67	- Get stuck in loops
68	- Need to be "euthanized" and restarted
69
70	With fungible agents, you never have to worry about that stuff.
71
72	When an agent dies:
73	1. The bead it was working on remains marked in-progress
74	2. Any other agent can pick it up
75	3. No special replacement logic needed
76	4. No dependency on that specific agent
77
78	---
79
80	## Scaling Properties
81
82	When you want to move faster:
83	- Simply spin up more agents at any time
84	- Don't need to think about whether there's an imbalance between the counts of each agent role
85	- 3 agents → 10 agents → 20 agents: just add more, they all do the same thing
86
87	Agent fungibility lets you go much faster and scales better, especially as:
88	- Projects get larger and more complex
89	- Agent counts increase past 10 working on the same project at the same time
90
91	---
92
93	## The Analogy: Fountain Codes
94
95	This is why I'm a huge fan of things like fountain codes (like RaptorQ) for file storage:
96
97	> Turn a file into an endless stream of fungible blobs with very little overhead and the user can catch any blob in any order and each new blob helps them reconstruct the file; there's no "rarest chunk" like with BitTorrent.
98
99	In software development:
100	- Each bead is a "blob"
101	- Any agent can work on any bead
102	- There's no "critical specialist" that becomes a bottleneck
103	- The system is resilient to partial failures
104
105	---
106
107	## When Specialized Roles DO Make Sense
108
109	There are contexts where designated roles for agents work better. Example: Automated scientific inquiry (like BrennerBot).
110
111	Why it works there:
112	- The discourse itself (the back and forth between different agent types) is the core mechanism of surfacing truth
113	- The debate structure IS the point
114
115	Why it doesn't work for software:
116	- In software development, we just want good code that works
117	- The goal is output, not discourse
118	- Fungibility removes bottlenecks and single points of failure
119
120	---
121
122	## The Clockwork Deity Philosophy
123
124	> Most of my agent tooling and workflows are about removing me from the equation, because there's only one of me, but a potentially unlimited number of agents.
125
126	The way to do this:
127	1. Front-load all input in planning phases — Use GPT Pro with Extended Reasoning, multiple iterations, blend feedback from all frontier models
128	2. Offload task structuring — Convert plans to beads with full d

People who liked this also viewed

6 picks

Type