Externalization, as proposed by Externalization in LLM Agents: A Unified Review of Memory, Skills, Protocols and Harness Engineering, is the progressive relocation of cognitive burdens from a model’s internal computation into persistent, inspectable and reusable external structures. Its authors present it as the transition logic of recent agent design — the mechanism that explains why each architectural shift has occurred and what form of reliability it sought to preserve — rather than as a description of any one system. The claim attached to it is about where reliable agency comes from: not from ever-larger models alone, but from restructuring task demands so that internal capabilities and external infrastructure jointly cover the competencies required.
Usage
The term is borrowed from Donald Norman’s theory of cognitive artifacts, and the borrowing carries its central point: an external aid does not merely amplify an unchanged internal ability, it often transforms the task. The paper’s examples are Norman’s — a shopping list turns a hard recall problem into a recognition problem; a map turns hidden spatial relations into visible structure — and the power of an artifact is therefore located in representational transformation rather than in added capacity.
Applied to language-model agents, the review identifies three dimensions of externalization, each named by the burden it relocates and the transformation it performs.
- Memory externalizes state across time. Rather than treating the context window as the sole carrier of history, accumulated knowledge — user preferences, prior trajectories, resolved ambiguities, domain facts — persists beyond a single session and is selectively retrieved. The transformation is from recall to recognition.
- Skills externalize procedural expertise. Rather than relying on weights to regenerate task-specific know-how on every invocation, procedures, best practices and operating guidance are packaged into reusable artifacts. The transformation is from generation to composition.
- Protocols externalize interaction structure. Rather than coordinating with tools, services and other agents through ad hoc prompt-level convention, explicit machine-readable contracts define discovery, invocation, delegation and permission management. The transformation is from ad hoc to structured.
The harness is the engineering layer that hosts all three and supplies the orchestration logic, constraints, observability and feedback loops that make externalized cognition cohere. The review is explicit that it is not a fourth kind of externalization but the runtime within which the other three operate.
The same paper extends the concept to parts of the runtime not usually described this way. It reads sandboxing as serving the same representational function as memory or skills — a cognitive boundary that changes what the model must reason about by removing irrelevant state — and reads permissions and policy configuration as externalized governance, constraints that would otherwise be embedded in prompts or enforced by post-hoc filtering, encoded instead as declarative rules the harness enforces at runtime.
When It Applies
- Applies where the difficulty is a mismatch rather than a capability gap. The review’s framing is that models are strong at flexible synthesis and reasoning over provided information and weaker at stable long-term memory, procedural repeatability and governed interaction with external systems; externalization is a response to that specific asymmetry.
- Applies most visibly to three recurrent problems the paper maps onto its three dimensions: a continuity problem arising from finite context and weak session memory, a variance problem arising from long procedures being rederived rather than executed consistently, and a coordination problem arising from brittle free-form interaction with tools and collaborators.
- Assumes the externalized structures can be kept coherent with each other. The review stresses that the dimensions do not evolve in isolation — memory expansion competes with skill loading for scarce context budget, protocol standardisation constrains how capabilities are packaged, and skill execution produces traces that later become memory — so a harness is needed to mediate them.
- Assumes the boundary between agent and environment is a design choice. The authors take this engineering insight from the distributed- and extended-cognition tradition while stating explicitly that they do not adopt that tradition’s stronger ontological claims.
- Is not a claim that weights stop mattering. The paper describes its three layers — Weights, Context, Harness — as layered rather than mutually exclusive, and says weights remain important even in the most infrastructure-heavy systems; what changes is where developers place the system’s mutable intelligence.
- Well grounded as a reading of current practice but presented as a synthesis rather than a measured result. The review is a systems-level argument organised around four claims about convergence, and it names measurement of externalization itself as an open problem.