Your AI Strategy Needs a Second Opinion. We Built a Free One.

I spend a lot of time watching enterprise AI decision-making up close. And there is a pattern I keep seeing that worries me more than any particular technology choice.

It goes like this. An executive hears about a new capability. They gather their leadership team. Someone senior states an opinion early, framing the question in a particular direction. The team nods. A few people offer supporting arguments. Someone raises a mild concern but doesn’t push it. A decision gets made. Everyone leaves feeling good about the process.

Six months later, the assumptions that never got tested surface as problems. The decision looked right in the room, but the room was never genuinely divided.

This is not a failure of talent. It is a structural failure of how groups make decisions. So I built a fix. It is called the Hermes Council. It is open source and free at github.com/magnus919/hermes-council.

What It Produced on Its First Real Question

I asked the council to debate the hardest enterprise AI question I could think of: should we build proprietary AI models or rely on third-party APIs for our core business functions?

Three agents were composed from scratch for this question alone. An ML infrastructure lead scarred by failed custom stacks. A platform economist who models total cost of ownership. An organizational learning theorist who studies how companies fail to capture what they learn.

They wrote independent failure histories before anyone stated a position. Then they debated each other. Then this happened.

The First Crack

Elena, the infrastructure lead, entered convinced that build was the right path. Her premortem scenario was specific and damning:

Eighteen months post-decision, the company is being acquired for pennies on the dollar by a competitor that took the opposite bet. The proprietary NLP stack we built absorbed three full squads for fourteen months, headcount that could have been shipping product features instead. The model we trained on our 2024 data could not adapt to the 2025 paradigm shift.

But when she read James’s position and pried into his reasoning, something shifted. She realized she had been thinking about lock-in wrong. The concession came in the cross-examination:

The switching cost question is symmetric in a way I had not fully articulated. API lock-in and custom-stack lock-in are both real. They just operate on different time horizons and have different exit costs.

The Second Crack

James, the economist, had built his position around the margin math of per-inference cost. But Elena’s argument about maintenance burden forced him to revise:

The carrying cost of bespoke maintenance, teams spending 60% of ML engineering time keeping custom infrastructure alive, reframes the build/buy calculation significantly.

The Third Crack

Priya, the organizational learning theorist, had been quiet through the early cross-examination. When she spoke, she identified a failure mode neither of the other two had modeled:

The organization optimized for engineering output over organizational learning. Built technically adequate models but systematically failed to capture the learning that would have told leadership when to stop.

What Emerged

The debate produced a framework that did not exist when the agents started. The build versus buy decision is not a binary choice. It is a three-axis tradeoff between scale economics, maintenance burden, and organizational learning capacity.

API lock-in compounds over months: pricing changes, deprecations, vendor strategy shifts. Custom-stack lock-in compounds over quarters: talent attrition, pipeline decay, architectural drift. The right choice depends on which time horizon your organization can actually manage. If you cannot sustain the learning cycle, the build path will produce technically adequate models that systematically fail, and you will not know until it is too late.

Elena entered the debate convinced that build was the right call, with a shelf full of reasons why. She left having conceded the central premise of her own argument to someone who started on the opposite side. That is not a failure of her reasoning. It is what structured debate is supposed to do.

This Is Not How Most Strategy Meetings Work

The research on group decision-making is clear and uncomfortable. Karadzhov et al. (2024) studied 500 group deliberation sessions. Diversity of initial positions among group members was a stronger predictor of performance gain than having a correct individual in the group. Probing for reasoning had a correlation of 0.41 with performance gain. Proposing solutions had a weaker effect.

Groups converge on solutions too quickly. The best performing teams use what Nesta’s collective intelligence review calls bursty communication: short, intense periods of structured disagreement separated by independent reflection.

Your leadership team almost certainly does the opposite. The structure of the meeting rewards agreement and punishes friction.

How It Works

The Hermes Council replaces unstructured discussion with structured debate through five phases:

• A premortem where each agent writes a failure history before anyone stakes a position
• Independent position formation that prevents anchoring to the first voice in the room
• Cross-examination where agents probe each other’s reasoning (this is where the insight lives)
• Assumption mapping where each agent identifies what would need to be true for opposing positions to be correct
• A synthesis that surfaces the decision landscape, not a forced recommendation

The most important design choice: the council never forces consensus. Forced consensus produces false consensus, agents agreeing on conclusions they do not believe. A decision landscape lets the person who actually has to make the call see the tension clearly.

Every agent reports confidence before and after the debate. If mean confidence drops and dispersion widens, the council surfaced genuine doubt. If it rises and narrows, that is the signature of groupthink. The council debated itself and mean confidence dropped from 0.80 to 0.70. It passed its own test.

Get It

The Hermes Council is open source, MIT licensed, and free. Go get it at github.com/magnus919/hermes-council. No signup, no API key, no vendor. Install it in under a minute if you already run Hermes Agent, an open source framework by Nous Research. One skill file, one orchestration script, zero new infrastructure.

Run it on a decision you are wrestling with right now. Not because it will give you a clean answer. Because it will give you a better map.