T28

The claim establishes a normative design principle: agents should be architecturally prevented from modifying configuration-tier (Tier 3) decisions, even though technically they can. This is contested by ML engineers who argue that adaptive agents that update their own configuration (ICP, messaging, objective function) based on outcome data are theoretically superior to fixed-configuration agents — the tier boundary prevents the system from learning at the strategy level. It is also contested by vendors who want to sell 'autonomous strategy optimization' capabilities. The counter-argument is that the strategy tier requires human judgment about objectives (T25), PMF assessment (T7), and ethical guardrails that agents are not currently equipped to manage.

In the GTM World Model, Tier 3 (Strategy) holds the configuration that constrains the agent: ICP definition, motion selection, messaging framework, do-not-contact rules, and objective function. These configuration choices require the type of judgment that integrates: market context (T0, T22), capital regime (T3), PMF assessment (T7, T21), and ethical constraints that cannot be fully encoded in a reward function. The tier boundary is normative because: (a) current agents lack the contextual model to safely update ICP or messaging without human oversight — wrong ICP updates could direct the agent to target sensitive buyer segments, competitors, or regulatory-constrained accounts; (b) the objective function (T25) requires explicit tradeoff choices between growth, efficiency, survival, and positioning that are fundamentally political decisions requiring human authority; (c) the human-in-the-loop data collection mechanism (T9) requires that humans retain authority over the configuration that shapes the learning signal — otherwise the agent is optimizing its own reward function without external calibration. 'Porous for humans' means the boundary is inspectable and overridable by human operators. 'Gated for agents' means the boundary is enforced architecturally: the agent cannot write to the configuration tier, only read from it.

• Documented agentic GTM failure modes: 11x and ZoomInfo AI SDR failures involved agents implicitly adapting their outreach patterns in ways that conflicted with the intended configuration — topology drift at the configuration level (T22)
• AI safety research (Amodei et al. 'Concrete Problems in AI Safety'): reward hacking and objective misspecification are most dangerous when the agent has write access to its own reward specification — the same principle applies to GTM configuration
• Salesforce Einstein AI governance research: enterprise customers required explicit human approval for any agent-initiated change to account routing rules, ICP scoring models, or messaging templates — confirming the normative value of the gated boundary in practice

• Some configuration parameters (scoring model weights, sequence timing) are legitimately optimizable by agents without human oversight — the boundary may be drawn too broadly
• Reinforcement learning agents that learn ICP segmentation from outcome data may discover more effective segments than human analysts, making the gated boundary a constraint on optimization quality
• The practical enforcement of the gated boundary requires architectural investment that most current GTM tech stacks have not made — the normative claim is correct but technically unrealized in most deployments

When implementing agentic GTM systems, explicitly architect the configuration-tier boundary: create a read-only configuration store that agents can access but not modify. Any agent action that would effectively modify configuration (routing all outreach to a new segment, changing messaging significantly, updating the objective scoring model) should require a human approval workflow before taking effect. Audit quarterly: run a log analysis of all agent actions over the past quarter and flag any that appear to be systematically changing who or how the agent targets. These are de-facto configuration changes that require explicit human review, even if no explicit configuration file was modified.