This review argues that the most consequential recent advances in LLM agents are not changes to models but changes to the runtime around them, and it supplies a single concept to explain why the separate literatures on memory, skills, protocols and harnesses are converging. That concept is externalization: the progressive relocation of cognitive burdens from the model’s internal computation into persistent, inspectable, reusable external structures.
The theoretical anchor is Donald Norman’s account of cognitive artifacts, from which the authors take the claim that an external aid does not simply amplify an unchanged internal ability but transforms the task itself. Their recurring illustrations are Norman’s: a shopping list does not expand biological memory capacity, it converts a difficult recall problem into a recognition problem; a map does not make navigation stronger, it converts hidden spatial relations into visible structure. The paper extends this into a deliberate parallel between the arc of human cognitive externalization — language, writing, printing, digital computation — and the arc it traces for LLM agents.
The review’s structure follows three externalization dimensions plus the layer that unifies them. Memory externalizes state across time, converting recall into recognition. Skills externalize procedural expertise, converting improvised generation into composition. Protocols externalize interaction structure, converting ad hoc coordination into governed contract. The harness is not a fourth kind of externalization but the runtime environment within which the three operate and interact.
Key Points
- The paper frames recent history as a progressive outward movement across three layers — Weights, Context, and Harness — and is careful that these are layered rather than mutually exclusive: weights remain important even in the most infrastructure-heavy systems, but each stage changes where developers place the system’s mutable intelligence and therefore where they invest engineering effort.
- It states the central limitation of weight-space capability as a governance problem rather than a capability one: parametric knowledge is difficult to selectively update, compose and govern, and auditing why a model behaved a certain way is hard because the relevant knowledge is distributed across billions of parameters rather than encoded as inspectable modules.
- It maps three recurrent mismatches of an unaided model onto the three externalization dimensions: a continuity problem (finite context and weak session memory) that memory addresses, a variance problem (long procedures rederived rather than executed consistently) that skills address, and a coordination problem (brittle interaction with tools and collaborators under free-form prompting) that protocols address.
- It reads the context-centric stage through the same Norman lens: the difficult question “does the model know fact X?” is converted into the easier one “given that fact X has been placed in context, can the model use it?” — and notes this mirrors the recall-to-recognition shift associated with writing in the human arc.
- It argues that expanding context windows does not dissolve the underlying tension, citing the “lost in the middle” finding that models attend unevenly across long inputs, and observing that context is also ephemeral: unless state is externalized elsewhere, every new session begins with partial amnesia.
- On the harness it makes a definitional claim rather than an implementational one: a harness is not an implementation convenience layered on a capable model but the designed cognitive environment within which externalized modules become jointly effective. Agency, on this account, is not located in the model alone but emerges from the coupling of the model with the environment that organises its cognition into action.
- It notes the term is still consolidating and presents its own characterisation as a synthesis of recurring patterns in current systems rather than a closed definition.
- It decomposes harness design into six analytical dimensions grouped under three operational surfaces — Permission, Control and Observability: agent loop and control flow, sandboxing and execution isolation, human oversight and approval gates, observability and structured feedback, configuration and policy encoding, and context budget management. None of the six is itself a form of externalization; they are the coordinative infrastructure that makes the three modules cohere.
- On control it argues that termination, recursion and cost limits are not secondary safety measures but definitional: they set the operational envelope within which reasoning unfolds, and a well-tuned loop makes an agent more reliable not by making the model smarter but by bounding the space of possible execution paths.
- It makes the same move for sandboxing, arguing an execution boundary is not merely a security fence but a cognitive boundary that simplifies the operating environment by removing irrelevant state and restricting dangerous actions — serving the same representational function as other forms of externalization.
- It treats autonomy as a configurable parameter of the harness rather than a binary property of the agent, adjustable per task, per tool and per organisational policy, and describes three common oversight patterns: pre-execution approval, post-execution review, and escalation triggers, with hook systems generalising the pattern by attaching arbitrary logic to lifecycle events.
- It argues observability is the mechanism by which a harness learns from its own operation, not an auxiliary convenience: without structured traces the feedback loops that connect execution outcomes back to the modules that produced them cannot operate, and the harness remains a static scaffold rather than an adaptive system.
- It frames permissions and policies as externalized governance — constraints that would otherwise be embedded in prompts or enforced by post-hoc filtering are instead declarative rules the harness enforces at runtime, stratified across user, project and organisation scopes so that the same base agent can operate under different policy regimes without changing the model or its skill artifacts.
- It treats context budget as a harness-level coordination problem that no single module can solve, because memory retrieval, skill loading, protocol schemas, tool descriptions and the model’s own reasoning traces all compete for the same finite allocation, and the optimal split depends on the current phase of execution.
- It reads the convergence of structurally similar harnesses across products with different lineages as analytically significant — evidence that the six dimensions are structural requirements of externalized agency rather than incidental implementation choices.
Notes
The authors draw on the tradition of distributed and extended cognition for its engineering insight — that the boundary between agent and environment is a design choice with real performance consequences — while stating explicitly that they do not commit to that tradition’s stronger ontological claims. They describe their focus as pragmatic, measuring the value of externalization by the reliability, composability and governability of the resulting system.
The review positions itself against existing surveys of retrieval-augmented generation, tool learning, agent architectures and protocol interoperability, arguing that what remains underdeveloped is not another component-level survey but a common account of why these developments are converging as forms of externalization.
It also emphasises that the dimensions do not evolve independently: memory expansion competes with skill loading for context budget, protocol standardisation improves interoperability while constraining how capabilities are packaged, and skill execution generates traces that later become memory.
See Externalization for the organising concept and Harness Engineering for the unifying layer.