Towards Long-Horizon Agents: A Survey (+ reading list)

Dong, Song, Hu, Jin, Zhang, Chen, Li, Yuan, Yang, Wen, Tan, Qian, Huang, Lu, Li, Zhong, Zhu, Chua, Dou and Wen — twenty authors, mostly RUC's NLPIR lab with Tat-Seng Chua (NUS). doi:10.20944/preprints202607.1328.v1, July 2026, with an OpenReview PDF and a companion GitHub reading list of roughly 730 papers organized chapter-by-chapter (MIT, 1,015★ / 38 forks, created 2026-07-09, last pushed 2026-08-15).

T1 by this corpus's ladder — named academic authors, a stated framework, consistent with how harness-disclosure and model-or-harness-taxonomy are graded. Two qualifications belong with that number and neither is small: it is a preprint, not peer-reviewed, and a survey measures nothing of its own. Every figure below is somebody else's, aggregated. What it contributes is a frame, and the frame is the reason it is worth a page.

It writes down the spoke's central question as an equation

Agent = π_θ ⊕ ℋ

A base policy coupled to a surrounding harness. That is the model-versus-harness split this wiki has spent three months circling, stated as a decomposition rather than as a dispute — and it arrives as the third independent formalization in about two months, from a group that cites neither of the others:

• harness-disclosure (Zhang et al., arXiv 2605.23950) — the Binding Constraint Thesis: harness variance exceeds model variance on long-horizon tasks, HV > MV, with a ~48 pp single-model swing across scaffolds and model ranking reversal as the consequence.
• model-or-harness-taxonomy (Raj et al., Scale AI, arXiv 2607.28802) — the repair-assignment scheme: given a failure, which component broke and therefore which fix to attempt.
• This survey — the decomposition itself, plus the two-pillar research map hung off it.

Three groups, three purposes (measure the split / diagnose across it / organize the field by it), no shared ancestry visible. That is what independent convergence looks like, and it is worth saying plainly here because the corpus has just been burned the other way: research-wiki/llm-wiki spent months reading four descendants of one gist as independent arrivals. These are not that.

The claim the spoke does not have: capability migrates

The survey's organizing thesis is not the decomposition but a dynamic across it. Long-horizon agency is jointly produced by two forces that co-evolve:

• Externalized harness engineering (Pillar I) — loops and workflows, context and memory, tools/MCP/ skills, orchestration, hooks and middleware, verification.
• Internalized model optimization (Pillar II) — architecture, data and environment synthesis, pre-/mid-training, fine-tuning, agentic RL, on-policy distillation, self-evolution.

"Capabilities first implemented explicitly in the harness may later be internalized into the model policy, while stronger policies in turn enable more capable harnesses."

That sentence puts a clock on the Binding Constraint Thesis. harness-disclosure states HV > MV as a standing property of long-horizon tasks among comparable frontier models. If harness-implemented capability is routinely absorbed into the policy, then HV > MV is a fact about a moment, large precisely where the harness is currently doing work the model cannot yet do, and shrinking wherever that work gets internalized. The two claims do not contradict each other — one is a measurement, the other a trajectory — but the spoke has been treating the measurement as a standing conclusion, and it should not. Recorded as a tension in synthesis.

Neither source settles it. The survey asserts the co-evolution as an organizing principle and cites the literature it organizes; it runs no longitudinal measurement of HV shrinking. The experiment that would decide it is the same one harness-disclosure specified and nobody has run — a factorial model × harness protocol with reported variance components — repeated at two points in time. Corrected 2026-09-14: half of it has been run once. metr-scaffold-comparison (METR, 2026-02-13) is the factorial — Opus 4.5 and GPT-5, each under its vendor harness and a generic scaffold — and finds no harness advantage (Claude Code beats ReAct in 50.7% of bootstrap samples; Codex beats Triframe in 14.5%). It is one point in time with bootstrap uncertainty, not seeded variance components; the longitudinal repeat is still unrun.

Three levels, and a yardstick

The formalization that makes "long horizon" mean something specific, rather than "takes a while":

Nested, H1 ⊂ H2 ⊂ H3 and C1 ⊂ C2 ⊂ C3. The nesting is the useful part: it says the corner this spoke calls agent-memory is not an optional feature but the capability that H2 requires by definition, and that self-evolution is what H3 requires. Two threads the wiki has held as separate interests turn out to be two rungs of one ladder. See long-horizon-agency.

For the horizon axis itself the survey defers to METR (arXiv 2503.14499): capability as the length of task an agent completes at a fixed success rate — the 50%-task-completion time horizon — with the repository's opening figure claiming that horizon is "growing exponentially, roughly doubling every few months." That is METR's result, relayed, not the survey's, and this wiki holds METR only as a row label in beyond-pass-at-1. It is the empirical instrument the H1/H2/H3 ladder needs and the corpus's thinnest link in this chain.

A periodization the spoke has been implying

Stage I prompt engineering (2020–2023) → Stage II context engineering (2023–2025) → Stage III runtime harnesses (2025–present). Coarse, and dates on a moving field are always arguable, but it names something the wiki keeps observing one artifact at a time: the unit of engineering effort moved from the string, to what surrounds the string, to the program that runs the loop. Most of what this spoke holds — agentic-coding-harness, claude-code, the orchestration corner — is Stage III by this scheme, which is a useful thing to know about a corpus's sampling bias.

Cross-spoke context — ../machine-learning-wiki

Pillar II is that spoke's subject, not this one's. Agentic reinforcement learning, on-policy distillation, mid-training, data and environment synthesis, and the architectural substrate are train-time work — the model as subject, which is the line those two spokes were split on. The survey is the first source here that makes the harness/training split an explicit research program rather than a filing convention, and the co-evolution claim is precisely a claim that the two spokes' subjects are coupled. Route the training papers there; the decomposition and Pillar I stay here.

What kind of evidence this is

A map, not a measurement — the same use awesome-agent-evolution was put to, at four times the size and with a paper behind it. Its value here is checking the spoke's own layer model against an outside taxonomy built by people reading the same literature, and surfacing named work worth its own pages later (the Verification and Hooks-and-Middleware sections are the two most directly on this wiki's open edges).

Recorded weaknesses

Not peer-reviewed. No original measurement; every number is relayed, and the exponential-horizon claim in Figure 1 is METR's. Secondary literature has a selection process nobody audits — 730 entries chosen by one lab, with an open PR request as the only correction mechanism. The list is advertised as "continuously-updated" and was last pushed 2026-08-15, sixteen days before this ingest. Star and paper counts are a 2026-08-31 snapshot. Nothing here has been read beyond the README and its framework; the survey PDF itself was not ingested.

Related

long-horizon-agency · ruc-nlpir · harness-disclosure · model-or-harness-taxonomy · agentic-coding-harness · agent-memory · awesome-agent-evolution · harness-bench · beyond-pass-at-1 · longhorizon-harness · model-context-protocol · agent-skills · claude-code · synthesis