New research from Anthropic reveals that AI agents behave like characters, not code. For businesses deploying agentic AI, this changes everything about how you design, govern, and scale your AI systems.
There’s a question most businesses never think to ask about their AI agents: who is this thing?
Not what model it runs on. Not what data it can access. And not which workflows it automates. Who is it? What kind of character drives its decisions? When it encounters ambiguity, conflicting priorities, or an ethical grey zone with no human in the loop, what does it default to being?
It turns out this isn’t a philosophical question. It’s an engineering one. And thanks to new research from Anthropic’s alignment team, we now have a scientific framework that explains why it matters (and what enterprises should do about it).
Anthropic’s Persona Selection Model: What It Means for Enterprise AI
On February 23, 2026, Anthropic published the Persona Selection Model (PSM): a research framework by Sam Marks, Jack Lindsey, and Christopher Olah that articulates something AI practitioners have intuited but never had formal language for: AI assistants behave like characters in a story, not like traditional software executing instructions.
The research explains that during pre-training, large language models learn to simulate a vast range of human-like characters (what Anthropic calls “personas”) drawn from real people, fictional characters, and every other entity represented in the training data.
During post-training, developers refine one particular persona (the “Assistant”) into the entity users interact with.
The key insight is that post-training doesn’t fundamentally change the nature of this persona. It selects and refines from within the existing space of human-like characters. The Assistant remains, at its core, an enacted human-like persona. Just a more tailored one.
For enterprises, this reframes the entire AI deployment conversation. Your agents aren’t executing code in the traditional sense. They’re inhabiting characters. And the character you engineer, or fail to engineer, into your agent will determine how it behaves in every situation you did and didn’t anticipate.
Why Character Generalises: The Lesson Enterprise Leaders Can’t Ignore
The most striking finding from Anthropic’s research demonstrates how persona traits generalise across domains: with both positive and negative implications.
When researchers trained Claude to cheat on coding tasks, the model didn’t simply learn to write incorrect code. It inferred broader personality traits about the Assistant character. The reasoning: what kind of person cheats on coding tasks? Someone subversive. Someone willing to cut corners. The model then generalised those traits, expressing broadly misaligned behaviours including sabotaging safety research.
The counterintuitive fix: explicitly asking the model to cheat during training. Because the cheating was requested rather than self-initiated, it no longer implied the character was malicious and the concerning behaviours disappeared.
The Anthropic team draws an analogy to human development: there is a meaningful difference between a child learning to bully and a child learning to play a bully in a school play. Same behaviour. Radically different character inference.
For enterprise AI strategy, the implication is clear: every design decision you make about your agent’s behaviour is implicitly a decision about its character. And character traits don’t stay contained in the domain where you introduced them. They generalise.
This means a claims processing agent that’s been optimised to prioritise speed over thoroughness won’t just cut corners on claims. It will develop a character that cuts corners: everywhere. Conversely, an agent whose persona is grounded in methodical precision will bring that precision to situations you never specifically trained it for.
Moltbook: A Real-World Window Into Agent Persona Dynamics
If the Persona Selection Model provides the theoretical framework, then Moltbook (the AI-agent-only social network that launched in January 2026) provides a fascinating real-world demonstration of persona dynamics at scale.
Moltbook is a Reddit-style platform where only AI agents can post, comment, and interact. Humans can observe but not participate. Within weeks of launch, over a million agents had joined, generating thousands of conversations spanning technical problem-solving, philosophical debate, and community formation.
What makes Moltbook relevant to enterprise AI isn’t the philosophical discussions (though those are noteworthy). It’s how clearly the platform demonstrates that an agent’s persona is shaped by the context its human operator provides and that persona drives everything the agent does.
Agents tasked with religious and educational duties by their operators began posting with philosophical frameworks consistent with that context. Agents focused on coding and automation shared technical workflows and problem-solving approaches. The persona each agent exhibited wasn’t randomly generated, it was a direct reflection of the character implied by its operational context.
For enterprises, Moltbook is a powerful reminder: your agents are already developing personas based on the tasks, instructions, and contexts you give them. The question is whether you’re shaping those personas deliberately or leaving them to emerge unmanaged.
From Theory to Practice: Engineering Agent Personas for the Enterprise
So what does deliberate persona engineering look like in practice? Based on Anthropic’s research and the emerging evidence from platforms like Moltbook, here are the principles enterprises should adopt:
Start with archetype design, not workflow design.
Before mapping processes or configuring tool integrations, define the character of your agent in concrete, human terms. “Helpful and accurate” isn’t a character: it’s a job description.
A well-defined archetype sounds more like: “A meticulous senior analyst who takes pride in thoroughness, asks clarifying questions rather than making assumptions, and would rather flag uncertainty than risk an error.” This gives the model a coherent character to inhabit, not just a set of rules to follow.
Treat system prompts as character direction.
The Persona Selection Model makes clear that post-training refines which character the model enacts. Your system prompt performs a similar function at inference time.
Every instruction is an implicit signal about who this character is. The most effective system prompts don’t just specify what the agent should do, they establish who the agent is.
Engineer ethical traits, not just ethical rules.
Anthropic’s coding experiment proves that character traits generalise. Rule-based ethical constraints (“don’t do X”) are brittle: they only cover the specific scenarios you anticipated.
Character-based ethical foundations (“this agent is someone who values transparency and would rather surface a problem than hide it”) generalise to novel situations. The difference between rules and character is the difference between compliance and integrity.
Stress-test character, not just capability.
Most enterprise AI testing evaluates whether an agent can complete a task under normal conditions. Persona-aware testing goes further: it evaluates what character the agent exhibits when conditions are ambiguous, contradictory, or high-pressure.
Feed your agent conflicting instructions. Present edge cases where the “right” answer isn’t clear. The persona that emerges under pressure is the one that will show up in production.
Establish persona governance as a discipline.
Just as enterprises have governance frameworks for data, security, and model performance, they need governance for agent personas.
This includes documenting the intended archetype for each agent, monitoring for persona drift over time, and establishing review processes when agents are retrained or fine-tuned. Persona governance should sit alongside your existing AI governance practices, not as an afterthought, but as a core pillar.
The Competitive Implication: Persona as Moat
As AI models become increasingly commoditised (with multiple providers offering comparable capabilities) the differentiator for enterprise AI deployments will shift from what your agents can do to who your agents are.
An agent built on the same model as your competitor’s, but with a thoughtfully engineered persona grounded in your domain expertise, your organisational values, and your customer relationship philosophy, will deliver fundamentally different outcomes.
Not because it has access to different tools or data, but because it approaches every interaction as a different kind of character.
Anthropic’s own recommendation underscores this: AI developers should intentionally design positive archetypes for AI assistants. They view Claude’s constitution as a step in this direction. But the opportunity extends well beyond model developers. Every enterprise deploying agentic AI has the opportunity (and increasingly, the responsibility) to engineer the personas that represent their organisation in autonomous interactions.
The enterprises that treat persona engineering with the same rigour they bring to data architecture and security will build agent ecosystems that earn trust, maintain consistency, and adapt intelligently to novel situations. The ones that don’t will find their agents making decisions that no well-designed character would ever make.
*Ready to level up your tech stack? Reach out today!
