Pre-meeting prep used to take hours. A CEO I worked with in 2026 needed strategy memos that matched the voice of his CMO, his CTO, his CFO, and his COO, depending on which audience was reviewing the doc. Writing in five voices is slow. Getting a voice wrong means the document gets rewritten, which burns another afternoon. The agent council pattern solves this with five named personas living inside a single Claude Code skill, each tuned to a specific executive's communication shape and decision-making frame.
This is the architecture. Three instances of the pattern in production. What worked. What breaks.
Positioning and counterfactual cost of delay, flagged.
Growth story reframed; market narrative load-bearing.
Technical coverage gap; architectural tradeoffs noted.
Comp band stretched; dilution modeled against plan.
Onboarding capacity tight; integration risk real.
The agent council pattern in one paragraph
The pattern is five named agent personas, each modeled on a different C-suite role, all orchestrated through a single Claude Code skill. The skill takes a topic or a decision prompt and routes it to the personas relevant for that topic, returning documents in the right voice on the first draft. Each persona has its own system prompt, its own reference library (memos, past decisions, voice examples), and its own output format. The orchestrator decides which personas to invoke, whether to invoke them in sequence or in parallel, and how to merge their perspectives into a final document.
The architecture comes from the agent council case study I built for an NDA client in Q1 2026. Five personas, zero external vendors, ten times faster on pre-meeting prep.
Instance 1: pre-meeting memos in five voices
The most common use is pre-meeting preparation. A weekly leadership sync where the CEO needs to walk in with a one-page memo that frames the decision from each executive's perspective before the conversation starts.
In the before state, the CEO's assistant would draft a single document summarizing the topic. The CMO would then rewrite the marketing framing. The CFO would annotate the financial section. The CTO would add technical notes. Two days of ping-pong to produce a doc that was useful for thirty minutes of meeting.
In the after state, the agent council produces a five-section memo where each section is already in the corresponding executive's voice. The CMO section uses the CMO's vocabulary, decision criteria, and typical framing. The CTO section reads the same way the CTO would write it. The CEO's prep time drops from two afternoons to about twenty minutes of review.
The deliverable is not a perfectly polished final doc. The deliverable is a draft that each executive would recognize as plausibly their own work. Reviewers catch a few nuances. Those become edits. The document ships.
Instance 2: strategy review before a board update
Second instance. Before a board update, the CEO wants to see the draft framed three ways: from the CFO lens (financial trajectory and cash runway), from the COO lens (operational risk and hiring plan), from the CMO lens (growth story and market positioning). Same underlying data. Three different framings.
The council handles this cleanly. The CEO gives the skill the board update topic and the three lenses. Three parallel persona drafts come back. The CEO picks the one that best matches the board's current anxieties and uses the other two as internal-only context for the conversation. Reviewing three perspectives before a meeting is faster than writing one from scratch, because the mental work of "what would the CFO emphasize here" is already cached in the persona's system prompt.
The deeper win shows up in the month-over-month pattern. The board meetings stop being ambushes. The CEO walks in with the framing that matches the board's current focus, and the council has already flagged the counter-arguments from the other two lenses.
“The deliverable is not a perfectly polished final doc. It is a draft each executive would recognize as plausibly their own work.
”
Instance 3: hiring decision frame
Third instance. A senior hire decision. The council produces a decision frame from four perspectives: CFO (comp band and dilution math), CTO (technical coverage and architectural gaps), COO (operational load and integration risk), CEO (strategic position and counterfactual cost of delay).
This is where the pattern gets interesting. The personas disagree. The CFO persona flags the comp band as stretched. The CTO persona says the candidate closes a gap that's been open for six months. The COO persona worries about onboarding capacity. The CEO persona weighs the opportunity cost of leaving the gap open another quarter.
The orchestrator does not try to resolve the disagreement. It surfaces it. The human decision-maker walks into the hiring conversation with the four arguments already articulated. That is the output. An explicit decision frame, not a recommendation. The recommendation is the human's job.
What the pattern tells us about voice-matching in agent output
Three instances, one mechanism. The agent council pattern works because voice is a stable function of persona context. When the context is dense enough (who the person is, how they talk, what they emphasize, what they skip), the model reproduces the voice with high fidelity.
The trick is context density. A persona with two sentences of description does not work. The CEO persona in the working implementation has about 400 lines of description: decision-making framework, communication patterns, sample memos, vocabulary notes, known blind spots, default framings. Same for each of the other four. That's roughly 2,000 lines of persona context before any topic prompt enters the skill.
Most failed agent-persona attempts I've seen are context-starved. A two-line "act like my CMO" prompt produces generic CMO voice. It reads like a SaaS landing page. The output embarrasses the human who would never talk that way. The fix is not more clever prompting. It's more context, specifically targeted at voice calibration.
I've written about the broader context of this role in the creative-tech operator playbook, because the agent council is one instance of a larger pattern: solo operators carrying more context than their role was traditionally priced for, using agents to hold the typing. The daily sequencing that feeds the council is covered in the morning routine I run before the first meeting.
How to spot when the council is ready to ship
Two tests. The first is the "reviewer surprise" test. Show a persona's draft to the executive it models. If their first reaction is "this is basically what I would have written, I'd just change three things," the calibration is there. If their reaction is "this sounds nothing like me," the persona context is too thin or miscalibrated.
The second test is the "intercept" test. Put a persona draft in a thread with the executive's previous writing samples. If a reader can't reliably tell which one is the persona draft, the council is ready. If the persona draft stands out as smoother, more generic, or more enthusiastic than the executive's real writing, you have calibration work left to do.
Both tests are cheap. Both should run before the council ships into a real decision workflow. A council that hasn't passed them will produce documents that the executive has to discard and rewrite, which costs more time than writing from scratch.
Where this pattern breaks
The pattern breaks in three specific situations.
The first is when the personas need to reach consensus in-model. The council is designed for parallel drafting, not debate. Trying to get the agents to argue and converge produces mush. The human has to be the arbiter. If you remove the human, the council produces compromise documents that nobody actually agrees with.
The second is when the topic is outside the persona's documented domain. A CFO persona asked about engineering architecture will produce a draft, but it will be thin and wrong-shaped. The orchestrator has to know which personas are valid for a given topic and skip the others. This requires explicit topic routing, not trust that the model will self-select.
The third is when voice calibration drifts. Executive voice is not static. People's communication patterns evolve with the company's stage. A council calibrated in Q1 will need re-calibration by Q3 if the executives are actively writing in new directions. The skill needs a voice drift check every quarter, using fresh writing samples, or the outputs start to sound like the executives' past selves rather than their current ones.
Frequently asked questions
How long does it take to build a council from scratch?
For five personas at production calibration, budget one to two weeks of elapsed time. The work is mostly context authoring: collecting writing samples, documenting decision frameworks, capturing vocabulary. The Claude Code skill authoring itself is a day or two. The calibration loop with each executive takes the rest.
Do the personas need to be real executives or can they be roles?
Real executives produce stronger calibration because you have writing samples and decision logs to draw from. Role-only personas work for generic outputs but degrade fast when the output needs to sound like a specific person. If you're building for a single founder, model on that founder's real execs. If you're building a template for others, use role archetypes with clear pattern documentation.
What models work best for the council?
Claude 3.5 Sonnet and newer handles persona instruction well, especially with long context windows. The council I built runs on Claude Code, which means it has access to the full repo and persona files without rehydration. GPT-4 class models can do it but tend to drift toward a generic executive voice faster. Test with the executive's own writing before committing.
Can I share a council across multiple clients?
Persona files are client-specific by design. The orchestration skill (the Claude Code code that routes topics to personas) can be reused across clients. The personas themselves should not be shared because they contain voice, decision frameworks, and often references to real internal terminology. Treat personas as client confidential material.
How much does running a council cost per month?
If you already have a Claude Code subscription, the marginal cost of the council is close to zero. The API costs for a few dozen persona invocations per week are negligible against the time saved. The upfront cost is the context authoring, not the runtime. I've written about the broader tooling baseline in the AI tooling stack post.
Sources and specifics
- Pattern grounded in a five-persona agent council built for an NDA client in Q1 2026. See the agent council case study for the public write-up.
- Persona context density observed to work: approximately 400 lines per persona across decision framework, communication patterns, sample documents, and vocabulary notes.
- Pre-meeting prep time observed to drop from roughly two afternoons to approximately twenty minutes of review on the original engagement.
- The pattern is a building block in the One-Person Studio OS, which packages the full orchestration library with persona templates and voice calibration checklists.
- Model compatibility observations based on Claude 3.5 Sonnet and Claude 3.7 Sonnet; earlier models tend to drift toward generic executive voice without more aggressive prompt engineering.