Agent Stack Org Design
Attribution: YourVisionYourCreation LLC — yourvisionyourcreation.com
Doctrine class: YVYC original — agentic category
Universal So-What
An agent stack is an organization, whether or not it is designed like
one. One human running multiple functions through multiple AI systems
has built an org chart by accident — and accidental org charts produce
the classic failures: two agents owning the same lane, no agent owning
a critical one, authority nobody assigned, and handoffs nobody
designed. Org design principles predate AI by a century because
coordination problems do; this doctrine applies them to the stack on
purpose.
Core Doctrine
1. Functions First, Agents Second
The design order is fixed:
| Step |
Question |
| 1 |
What FUNCTIONS does this operation require? (strategy, engineering, finance, brand, legal, operations, comms...) |
| 2 |
Which functions does the HUMAN hold personally, and which delegate to the stack? |
| 3 |
Which agent/surface owns each delegated function? |
| 4 |
Where do functions hand off to each other, and how? |
Assembling agents first and discovering their functions later is how
stacks accrete into overlap and orphanage. The function map is the
org chart; agents fill boxes that already exist.
2. The Lane Charter — One Owner Per Function
Every agent in the stack carries a written charter:
- Owns: the functions this agent is the authority for
- Contributes: where it assists another agent's lane
- Never touches: the lanes explicitly outside its scope
- One function, one owning agent. Two agents owning one lane
produces contradictory output the human must referee — the
referee work silently consumes the time the stack was built to
save.
- Lane violations are protocol failures even when the output is
good: an agent freelancing in another's lane on Tuesday is the
contradiction machine of Thursday
3. Authority Mapping — Decide, Recommend, Execute
Every function marks where authority sits, using three levels:
- Execute: the agent acts without approval (drafts, research,
analysis, reversible work)
- Recommend: the agent proposes; the human decides (strategy
shifts, spending, commitments)
- Decide: reserved to the human alone (public exposure, legal,
financial, anything irreversible or reputation-bearing)
- The default for a new, unmapped situation is RECOMMEND — an agent
guessing at its own authority guesses upward over time
- Authority levels are per-function, not per-agent: the same agent
may execute research freely and only recommend on publication
4. Handoff Engineering
The seams between agents are where stacks fail:
- Every recurring handoff is designed: what travels (context,
files, decisions made), in what form, with what fidelity check
- Handoffs are truth-first: the receiving agent gets the actual
state — including what is broken, unfinished, or uncertain —
not a cleaned-up version that costs the next lane a discovery
phase
- The human is not the default transport layer: where the human
hand-carries every context packet between agents, the stack has
outsourced its coordination cost to its most expensive node —
document the handoff so it can eventually be systematized
- Cross-lane artifacts (a handoff document, a shared state file,
a control-plane repo) outrank verbal relay
5. The Single-Human Node Rules
For operations where one human runs the whole stack:
- The human's irreplaceable functions are named: judgment, priority,
relationships, authorization — the stack augments these, never
absorbs them
- Rotation across functions is a legitimate operating method, and
the stack is designed FOR it: every lane holds state well enough
that the human can leave and return without archaeology —
standing documents, decision logs, and next-actions per lane
- The bus test still applies at scale-of-one: critical state living
only in the human's head or one conversation thread is fragility
wearing efficiency's uniform — the stack's documents are the
succession plan
- Stack capacity is budgeted like any resource: every agent added
costs human attention to direct and verify; agents that cost more
direction than they return get consolidated or cut
6. The Stack Review Cadence
- The org design is versioned and reviewed on a cadence: functions
added or retired, lanes redrawn, authority promoted or demoted
based on demonstrated reliability
- Promotion is earned by record: an agent moves from recommend to
execute on a function after a documented run of sound
recommendations — trust is calibrated, not assumed (the same law
as trust-calibration in this library's foundation tier)
- Every stack failure (contradiction, orphaned task, authority
ambiguity) is traced to the org design question it exposes: the
incident is the audit
Common Failure Modes
| Failure |
Cause |
Correction |
| Two agents produce contradictory work |
Overlapping lanes, no charters |
One function, one owner — written |
| Critical function silently unowned |
Agents assembled before functions mapped |
Functions first; agents fill existing boxes |
| Agent acts beyond its authority |
Authority never mapped; default drifted upward |
Decide/recommend/execute per function; unmapped defaults to recommend |
| Every handoff loses context |
Human as transport layer, verbal relay |
Engineered handoffs; truth-first artifacts |
| Return from rotation requires archaeology |
Lanes hold no state |
Standing docs, decision logs, next-actions per lane |
| Stack grows, output does not |
Agent accretion without capacity budget |
Direction cost audited; consolidate or cut |
Non-Negotiables
- Functions are mapped before agents are assigned.
- Every lane has exactly one owning agent, chartered in writing.
- Authority is mapped per function; unmapped situations default to
recommend.
- Irreversible and reputation-bearing decisions are reserved to the
human.
- Handoffs are truth-first and engineered, not improvised.
- The stack design is versioned, reviewed, and promoted by record.
YourVisionYourCreation LLC — yourvisionyourcreation.com
Licensed under CC BY 4.0
1---2name: agent-stack-org-design3description: Activate whenever the task involves designing how multiple AI agents, tools, and surfaces map to human roles and organizational functions — one person augmented by an agent stack, a team distributing work across AI systems, deciding which agent owns which lane, where authority lives and where it delegates, or "how should I structure my AI setup." Trigger on multi-agent architecture, agent role definition, human-AI org charts, and any operation consolidating multiple functions into fewer humans via AI augmentation. Fire at design time — agent stacks assembled by accretion develop overlapping lanes, orphaned functions, and authority ambiguity that surfaces as chaos precisely when the stakes are highest.4license: CC BY 4.05---67# Agent Stack Org Design89**Attribution:** YourVisionYourCreation LLC — yourvisionyourcreation.com10**Doctrine class:** YVYC original — agentic category1112---1314## Universal So-What1516An agent stack is an organization, whether or not it is designed like17one. One human running multiple functions through multiple AI systems18has built an org chart by accident — and accidental org charts produce19the classic failures: two agents owning the same lane, no agent owning20a critical one, authority nobody assigned, and handoffs nobody21designed. Org design principles predate AI by a century because22coordination problems do; this doctrine applies them to the stack on23purpose.2425---2627## Core Doctrine2829### 1. Functions First, Agents Second3031The design order is fixed:3233| Step | Question |34|---|---|35| 1 | What FUNCTIONS does this operation require? (strategy, engineering, finance, brand, legal, operations, comms...) |36| 2 | Which functions does the HUMAN hold personally, and which delegate to the stack? |37| 3 | Which agent/surface owns each delegated function? |38| 4 | Where do functions hand off to each other, and how? |3940Assembling agents first and discovering their functions later is how41stacks accrete into overlap and orphanage. The function map is the42org chart; agents fill boxes that already exist.4344### 2. The Lane Charter — One Owner Per Function4546Every agent in the stack carries a written charter:4748- **Owns:** the functions this agent is the authority for49- **Contributes:** where it assists another agent's lane50- **Never touches:** the lanes explicitly outside its scope51- One function, one owning agent. Two agents owning one lane52 produces contradictory output the human must referee — the53 referee work silently consumes the time the stack was built to54 save.55- Lane violations are protocol failures even when the output is56 good: an agent freelancing in another's lane on Tuesday is the57 contradiction machine of Thursday5859### 3. Authority Mapping — Decide, Recommend, Execute6061Every function marks where authority sits, using three levels:6263- **Execute:** the agent acts without approval (drafts, research,64 analysis, reversible work)65- **Recommend:** the agent proposes; the human decides (strategy66 shifts, spending, commitments)67- **Decide:** reserved to the human alone (public exposure, legal,68 financial, anything irreversible or reputation-bearing)69- The default for a new, unmapped situation is RECOMMEND — an agent70 guessing at its own authority guesses upward over time71- Authority levels are per-function, not per-agent: the same agent72 may execute research freely and only recommend on publication7374### 4. Handoff Engineering7576The seams between agents are where stacks fail:7778- Every recurring handoff is designed: what travels (context,79 files, decisions made), in what form, with what fidelity check80- Handoffs are truth-first: the receiving agent gets the actual81 state — including what is broken, unfinished, or uncertain —82 not a cleaned-up version that costs the next lane a discovery83 phase84- The human is not the default transport layer: where the human85 hand-carries every context packet between agents, the stack has86 outsourced its coordination cost to its most expensive node —87 document the handoff so it can eventually be systematized88- Cross-lane artifacts (a handoff document, a shared state file,89 a control-plane repo) outrank verbal relay9091### 5. The Single-Human Node Rules9293For operations where one human runs the whole stack:9495- The human's irreplaceable functions are named: judgment, priority,96 relationships, authorization — the stack augments these, never97 absorbs them98- Rotation across functions is a legitimate operating method, and99 the stack is designed FOR it: every lane holds state well enough100 that the human can leave and return without archaeology —101 standing documents, decision logs, and next-actions per lane102- The bus test still applies at scale-of-one: critical state living103 only in the human's head or one conversation thread is fragility104 wearing efficiency's uniform — the stack's documents are the105 succession plan106- Stack capacity is budgeted like any resource: every agent added107 costs human attention to direct and verify; agents that cost more108 direction than they return get consolidated or cut109110### 6. The Stack Review Cadence111112- The org design is versioned and reviewed on a cadence: functions113 added or retired, lanes redrawn, authority promoted or demoted114 based on demonstrated reliability115- Promotion is earned by record: an agent moves from recommend to116 execute on a function after a documented run of sound117 recommendations — trust is calibrated, not assumed (the same law118 as trust-calibration in this library's foundation tier)119- Every stack failure (contradiction, orphaned task, authority120 ambiguity) is traced to the org design question it exposes: the121 incident is the audit122123---124125## Common Failure Modes126127| Failure | Cause | Correction |128|---|---|---|129| Two agents produce contradictory work | Overlapping lanes, no charters | One function, one owner — written |130| Critical function silently unowned | Agents assembled before functions mapped | Functions first; agents fill existing boxes |131| Agent acts beyond its authority | Authority never mapped; default drifted upward | Decide/recommend/execute per function; unmapped defaults to recommend |132| Every handoff loses context | Human as transport layer, verbal relay | Engineered handoffs; truth-first artifacts |133| Return from rotation requires archaeology | Lanes hold no state | Standing docs, decision logs, next-actions per lane |134| Stack grows, output does not | Agent accretion without capacity budget | Direction cost audited; consolidate or cut |135136---137138## Non-Negotiables1391401. Functions are mapped before agents are assigned.1412. Every lane has exactly one owning agent, chartered in writing.1423. Authority is mapped per function; unmapped situations default to143 recommend.1444. Irreversible and reputation-bearing decisions are reserved to the145 human.1465. Handoffs are truth-first and engineered, not improvised.1476. The stack design is versioned, reviewed, and promoted by record.148149---150151*YourVisionYourCreation LLC — yourvisionyourcreation.com*152*Licensed under CC BY 4.0*
Forexgod21