01 / Overview
Important decisions rarely fail because there was no argument for the option someone chose.
They fail because an assumption went unchallenged, an alternative was never considered, or disagreement disappeared too early.
Claude Council is an experiment in structuring AI reasoning around that problem.
Instead of asking one model:
"What should I do?"
the system creates a small decision process.
Five advisors analyze the same decision independently from different reasoning perspectives. Their responses are anonymized and reviewed by other agents, unusually strong consensus can trigger an adversarial debate, and two separate chairmen examine the majority and minority positions before producing the final recommendation.
The goal is not to make an AI "choose for you."
The goal is to build a process that makes weak assumptions, uncertainty, disagreement, and overlooked alternatives harder to hide.
02 / My Contribution
I designed Claude Council as a structured decision pipeline rather than a collection of personas responding to the same prompt.
What I built
- Designed the end-to-end multi-agent council architecture
- Created five advisors with deliberately different reasoning methods
- Implemented parallel advisor fan-out to preserve independent analysis
- Added a bias audit before deliberation begins
- Built anonymized peer review to reduce persona and presentation bias
- Designed consensus scoring and conditional adversarial debate
- Created separate consensus and dissent chairman passes
- Added explicit dissent preservation to the final recommendation
- Built adaptive Quick, Standard, and Deep execution modes
- Added confidence, assumptions, unknowns, and decision-breaker tracking
- Implemented persistent decision journaling and outcome recording
- Added historical lookup so previous decisions can inform future councils
- Built structured HTML reports and complete Markdown transcripts
The main problem I was exploring was not how to generate more opinions.
It was how to design interaction between agents so that additional reasoning actually adds useful disagreement instead of repeating the same answer several times.
03 / Architecture
Claude Council runs decisions through a staged deliberation pipeline.
Decision flow
| Stage | Responsibility |
|---|---|
| Triage | Decide whether the question actually deserves a council |
| History | Look for previous related decisions and outcomes |
| Framing | Structure the decision, stakes, options, and context |
| Bias Audit | Flag potential cognitive distortions before deliberation |
| Advisors | Generate independent analyses from different reasoning lenses |
| Peer Review | Evaluate advisor arguments without revealing their identities |
| Debate | Attack unusually strong consensus |
| Chairmen | Independently synthesize majority and minority interpretations |
| Dissent Pass | Restore valuable arguments that consensus may have softened |
| Journal | Record the decision for future calibration |
This separation is intentional.
An advisor cannot simply become the final answer.
Its reasoning has to survive review, comparison, disagreement, and synthesis first.
04 / The Council
The five advisors are designed around different failure modes in decision-making, rather than different writing personalities.
| Advisor | Question it pressures |
|---|---|
| Red Team | How could this fail? |
| First Principles | Which assumptions are we treating as facts? |
| Expansionist | What options are missing entirely? |
| Outsider | What would another discipline notice here? |
| Executor | Can this actually be executed with the available resources? |
Red Team
Works backwards from failure.
Instead of asking why an option might succeed, it assumes the decision went badly and attempts to identify:
- failure modes
- root causes
- missed signals
- safer alternatives
First Principles
Breaks the decision apart and challenges its starting assumptions.
It tries to identify when the council is reasoning inside a frame that may itself be wrong.
existing assumption
↓
break it apart
↓
reframe the problem
↓
compare new possibilities
Expansionist
Looks beyond the options explicitly provided by the user.
A decision framed as:
A vs B
may actually contain:
A vs B vs C vs hybrid vs delay vs experiment first
The Expansionist exists to prevent the council from optimizing inside an unnecessarily narrow option set.
Outsider
Approaches the problem through methods borrowed from another field.
The objective is to introduce reasoning patterns the other advisors may not naturally reach.
Executor
Pressure-tests feasibility.
It focuses on:
- available resources
- implementation difficulty
- reversibility
- execution sequence
- missing data
- expected impact
This acts as a counterweight to recommendations that are strategically attractive but operationally unrealistic.
05 / Engineering Decisions
Independence before collaboration
One of the most important architectural decisions was preventing the advisors from seeing each other's reasoning during the first pass.
Decision
┌────────┬────────┬────────┬────────┬────────┐
↓ ↓ ↓ ↓ ↓
Red Principles Expansion Outsider Executor
Team
│ │ │ │ │
└────────┴────────┴────────┴────────┴────────┘
↓
Comparison
If agents collaborate too early, they can anchor on whichever argument appears first.
Parallel fan-out gives each reasoning lens an opportunity to form its own conclusion before social influence enters the pipeline.
The council combines the analyses after independence has been established.
Anonymized peer review
Another design decision was separating the argument from the identity of the advisor that produced it.
Before peer review, advisor responses are mapped to anonymous labels:
Red Team ─┐
First Principles ─┤
Expansionist ─┼──→ A · B · C · D · E
Outsider ─┤
Executor ─┘
Persona-specific structural clues are also normalized where possible.
Reviewers therefore evaluate the reasoning itself instead of knowing:
"This is the Red Team argument, so naturally it is pessimistic."
Only after peer review are the original identities restored for final synthesis.
Confidence as structured output
Every advisor is required to expose more than a recommendation.
Each analysis includes:
CONFIDENCE
confidence:
assumptions:
what would change my mind:
unknowns:
This was important because confidence without context is not very useful.
An advisor saying:
confidence: high
matters much less than knowing:
confidence: high, assuming customer acquisition remains below X
The final council can therefore surface not only what it believes, but what must remain true for that recommendation to hold.
Adaptive depth instead of maximum reasoning every time
Not every decision deserves the same computational process.
Claude Council therefore supports three deliberation modes.
| Mode | Advisors | Peer Review | Debate | Best For |
|---|---|---|---|---|
| Quick | 3 | — | — | Low-cost, reversible decisions |
| Standard | 5 | ✓ | Conditional | Normal high-stakes decisions |
| Deep | 5 expanded | ✓ | Always | High-cost or difficult-to-reverse decisions |
The architecture trades additional computation for additional scrutiny only when the decision justifies it.
06 / The Challenge
Consensus is not always a success condition
A surprisingly difficult problem in multi-agent systems is that adding agents does not guarantee independent thinking.
Five agents can still converge on the same attractive argument.
That creates a dangerous failure mode:
agreement
≠
correctness
Claude Council therefore measures the strength of agreement during peer review.
If consensus is unusually strong, the system becomes more skeptical, not less.
A Prosecutor is asked to attack the dominant recommendation.
Only after that attack does a Defender receive both the consensus argument and the prosecution and attempt to defend it.
The debate therefore becomes sequential:
consensus → attack → defense
rather than two agents independently producing opposing essays.
Preserving useful disagreement
A second problem appears during synthesis.
Even if the council produced useful minority arguments, a final summarization step can accidentally erase them while creating a clean recommendation.
I addressed this with two separate chairman passes.
Council State
│
┌─────────┴─────────┐
↓ ↓
Consensus Chairman Dissent Chairman
│ │
└─────────┬─────────┘
↓
Dissent Preservation
↓
Final Verdict
The first chairman is intentionally majority-oriented.
The second is intentionally minority-oriented.
A final dissent-preservation pass compares both outputs and restores important arguments that the majority synthesis weakened or omitted.
The final goal is not artificial compromise.
It is:
make a recommendation while remaining explicit about why it could still be wrong.
07 / Decision Memory
Claude Council also keeps a persistent record of previous decisions.
Each completed run writes structured metadata to a journal, including:
- decision identifier
- execution mode
- advisor confidence
- chairman confidence
- flagged biases
- final recommendation
- preserved dissent
- generated report paths
- eventual outcome
A user can later record what actually happened:
decision
↓
recommendation
↓
real-world action
↓
outcome
↓
future council context
Before a new council begins, the system can search this history for related decisions.
That means previous decisions are not only archived.
Their outcomes can become context for future deliberation.
Closing the feedback loop
This creates a simple but important feedback cycle:
After enough recorded decisions, the system can perform meta-analysis over the journal to look for recurring calibration patterns.
For example:
- an advisor consistently overestimating risk
- one reasoning lens performing especially well for certain decisions
- repeated confidence mismatches
- assumptions that frequently turn out to be wrong
This moves the project from a stateless prompting workflow toward an experiment in decision-process calibration over time.
08 / Output
Every full council produces two artifacts.
Decision Report
A visual HTML report containing the important decision state:
- council recommendation
- confidence level
- dominant assumptions
- decision breakers
- bias flags
- advisor analyses
- agreement grid
- feasibility scoring
- peer reviews
- debate transcript
- dissent ledger
Full Transcript
A Markdown transcript preserving the complete deliberation process.
decision
↓
advisor reasoning
↓
peer reviews
↓
debate
↓
chairman synthesis
↓
dissent ledger
↓
final verdict
The report is intended for quickly understanding the conclusion.
The transcript exists for inspecting how the system reached it.
09 / Current State
Claude Council currently supports the complete decision workflow from question framing to persistent outcome tracking.
Working now
- Decision triage
- Prior-decision lookup
- Cognitive bias audit
- Quick / Standard / Deep execution modes
- Parallel multi-agent analysis
- Five specialized reasoning lenses
- Structured confidence reporting
- Advisor anonymization
- Parallel peer review
- Consensus-strength scoring
- Conditional adversarial debate
- Dual-chairman synthesis
- Minority dissent preservation
- Decision-science analysis in Deep mode
- Automatic confidence-based escalation
- HTML decision reports
- Markdown deliberation transcripts
- Persistent decision journal
- Outcome recording
- Historical meta-analysis
The complete system can be invoked directly inside Claude Code:
/claude-council <decision>
10 / What This Project Demonstrates
Claude Council started from a simple question:
Would several reasoning agents produce better decisions than one?
The more interesting engineering problem became determining how those agents should interact.
The project pushed me to think about:
- independence before collaboration
- orchestration instead of prompt chaining
- cognitive diversity instead of duplicated agents
- adversarial testing instead of trusting consensus
- explicit uncertainty instead of confident prose
- minority preservation instead of majority summarization
- adaptive compute instead of maximum complexity
- persistent outcomes instead of stateless conversations
- evaluation of the reasoning process rather than only the final answer
The value of the system does not come from having five agents instead of one.
It comes from giving those agents different responsibilities, controlled information flow, and opportunities to challenge one another.
11 / Next
There are several directions I want to explore further.
- Quantitatively evaluate council decisions against single-agent baselines
- Measure whether anonymization meaningfully changes reviewer rankings
- Compare debate strategies and consensus thresholds
- Improve calibration using recorded real-world outcomes
- Visualize advisor agreement and disagreement as a decision graph
- Track which assumptions most frequently cause recommendations to fail
- Experiment with dynamically selecting advisors based on decision type
- Evaluate whether more agents actually improve decisions or simply increase redundancy
- Build stronger tests for correlated reasoning between agents
The larger goal is to explore how multi-agent systems can produce better-calibrated reasoning, not simply more reasoning.