Dual‑Engine Evolution: A Systematic Survey of Long‑Horizon Agents

The paper "Towards Long‑Horizon Agents: A Survey – Foundation, Evolution, Harness, Optimization, Application, and Frontier" (jointly authored by researchers from Renmin, Peking, and Tsinghua universities) offers the first unified formal definition of a long‑horizon agent: Agent = πθ ⊕ H , where the base policy πθ is coupled with a runtime harness H that decides when the model is invoked, validates actions, and persists state.

Concept Clarification

Terms such as “self‑evolving Agent”, “autonomous Agent”, and “long‑running Agent” have been used interchangeably, obscuring the source of progress. The survey argues that long‑horizon capability is not an intrinsic property of the model alone but the product of external harness engineering (loops, memory, tools, orchestration, hooks, verification) co‑evolving with internal model optimization (architecture, data, training, RL, self‑evolution).

Definition and Taxonomy

Long‑horizon tasks are defined by internal logical dependencies rather than elapsed time: a task requires a coherent chain of inter‑dependent decisions to complete. The authors model this as a coupling of πθ and H, operating in a partially observable Markov decision process (POMDP) with sparse, delayed rewards, creating a credit‑assignment problem.

Three nested task levels (H1‑H3) and corresponding capability levels (C1‑C3) are introduced:

H1 / C1 : Context‑bound (minutes) tasks with tightly coupled decisions and interactive reasoning.

H2 / C2 : Cross‑context (hours‑to‑days) tasks requiring memory compression, read/write, and checkpoint‑style continuation.

H3 / C3 : Cross‑task (lifelong) tasks demanding skill accumulation and continual improvement.

Each level nests the previous one, adding new failure modes.

Empirical Growth

Using METR data up to 8 May 2026, the median task‑completion horizon for frontier agents grew from seconds (GPT‑2 era) to ~12 hours for Claude Opus 4.6 and ≥16 hours for Claude Mythos Preview—a four‑order‑of‑magnitude increase. The fitted doubling time is 196.5 days overall, shortened to 130.8 days for methods introduced after 2023, indicating accelerating progress.

Evolution Stages

Stage I – Prompt Engineering (2020‑2023) : Language prompts are the sole control surface; techniques such as CoT, Self‑Consistency, and Plan‑and‑Solve create early agentic behavior; InstructGPT marks the first internalisation of harness into policy.

Stage II – Context Engineering (2023‑2025) : Control expands to the information visible to the model; RAG, Toolformer, Gorilla, ReAct, and MemGPT introduce external evidence, tool use, and memory management.

Stage III – Runtime Harness (2025‑present) : Control shifts to the entire execution trajectory; loops, hooks, and governance mechanisms surround model operation, enabling true long‑horizon agents.

Harness – Six Core Components

Loops & Workflows : Linear, plan‑execute, and branch flows (maker‑checker) for explicit trajectory tracking.

Context & Memory : Context discard/compression/selection and persistent memory operations; the harness acts as the interface between workflow and state.

Tools, MCP & Skills : Tool APIs, proactive tool discovery, and skill libraries; MCP solves M×N integration, while skill libraries store reusable expertise.

Orchestration : Task decomposition, role assignment, topology coordination, and inter‑agent protocols (e.g., A2A).

Hooks & Middleware : Predefined, custom, and adaptive hooks driven by uncertainty, behavior, or threat, turning passive execution shells into active controllers.

Verification : Evaluation of correctness, safety, fidelity, and multi‑agent interactions at step‑level or episode‑level, using automatic labeling, reward shaping, and optimal verification strategies.

Optimization – Seven‑Stage Agentic Training Pipeline

Architecture Base : Explicit context, compressed state, hybrid architecture, high‑throughput mechanisms to represent, train, and decode long trajectories.

Data & Environment Synthesis : Generation of tasks, environments, and trajectories to provide verifiable long‑horizon experience.

Pre‑training / Mid‑training : Injection of reasoning priors, long‑context state, and multimodal perception so that agent behavior emerges natively.

Fine‑tuning : Instruction selection, curriculum learning, and distillation to instill disciplined behavior and tool usage.

Agentic RL : Addresses four key questions under sparse, delayed reward: credit assignment, policy optimisation, sampling strategy, and interaction mode.

On‑Policy Distillation : Consolidates behavior on the policy’s own state distribution, reducing distribution shift.

Self‑Evolution : Offline (generate‑filter‑train), online (learn during interaction), and co‑evolution with the environment to maintain a learnable capability boundary.

Applications – Five Interface‑Driven Forms

Applications are organised by the agent‑environment interface rather than industry domain, because the interface determines observation, action, state, and feedback shapes, and thus the long‑horizon difficulty. The five forms progress from bounded, verifiable environments to open, heterogeneous worlds, with corresponding benchmark suites (SWE‑bench, WebArena, OSWorld, GAIA, Terminal‑Bench, METR) listed in Table 7.

Frontier Challenges

The survey groups open problems into four axes and nine concrete directions (Table 8). A recurring judgement is that the next major breakthroughs will occur at the Harness layer.

Towards Long-Horizon Agents: A Survey Foundation, Evolution, Harness, Optimization, Application, and Frontier
https://github.com/RUC-NLPIR/Awesome-Long-Horizon-Agents
https://openreview.net/pdf?id=HyhfhlbWGh