Long-horizon agency

The capability an agent needs to keep working on one thing across many interdependent steps — "persistent iteration across reasoning, tool use, observation, and revision" — where the horizon runs from a task inside a single context window to a task stream with no end. Named and formalized in long-horizon-agents-survey (Dong et al., 2026), which is where everything on this page comes from.

It is the property most of this spoke has been about without having a word for. A harness exists to extend it; agent-memory exists to carry it across a boundary; harness-bench, ddr-bench and scaffold-effects-gaia measure where it fails.

The decomposition

Agent = π_θ ⊕ ℋ

A base policy coupled to a surrounding harness. Long-horizon agency is a property of the pair, not of either half — which is why the spoke's recurring model-versus-harness argument (harness-disclosure, model-or-harness-taxonomy) is better read as a question about where a given capability currently sits than about which half matters.

Three nested levels

H1 ⊂ H2 ⊂ H3, C1 ⊂ C2 ⊂ C3. The nesting carries the content. Memory is not a feature at H2; it is the definition of H2 — a task that outlives its context window cannot be done without cross-context state, so agent-memory and the memory-as-infrastructure thread stop being an interest of this wiki and become a structural requirement. The same move puts self-evolution at H3: accumulating experience across unrelated tasks is what an open-ended stream demands.

What "horizon" is measured in. METR's yardstick (arXiv 2503.14499): the length of task an agent completes at a fixed success rate — the 50%-task-completion time horizon. It is the only measured instrument in this frame, it is not this spoke's own, and the wiki holds it only in passing (beyond-pass-at-1). Treat the "doubling every few months" figure as METR's claim relayed, not as something checked here.

The two pillars, and why the distinction is load-bearing

Capability can be built outside the model or inside it:

• Externalized (harness) — loops and workflows, context and memory, tools/MCP/ skills, orchestration, hooks and middleware, verification.
• Internalized (model) — architecture, data and environment synthesis, pre-/mid-training, fine-tuning, agentic RL, on-policy distillation, self-evolution. This half is ../machine-learning-wiki's subject.

The survey's claim is that these co-evolve: a capability first bolted on in the harness may later be trained into the policy, and a stronger policy then supports a more capable harness.

That makes the placement of any given capability temporary, which has a consequence for how this wiki reads its own evidence. harness-disclosure measured harness variance exceeding model variance and stated it as a thesis. Under the co-evolution account that gap is largest exactly where the harness is compensating for something the model cannot yet do, and it closes wherever the compensation gets trained in. So a harness-effect measurement is dated by construction. Nobody has run the longitudinal version that would show the gap moving; see synthesis. (Checked 2026-09-14: the single-time-point factorial now exists — metr-scaffold-comparison — and the gap it measures at the frontier is about zero, which is what the co-evolution account predicts for the strongest models; the repeat over time is still missing.)

Three stages of where the effort went

Prompt engineering (2020–2023) → context engineering (2023–2025) → runtime harnesses (2025–present). The unit of engineering moved from the string, to what surrounds the string, to the program running the loop. Coarse and arguable at the boundaries, but it dates this spoke's own corpus: nearly everything here is Stage III.

Related

long-horizon-agents-survey · agentic-coding-harness · agent-memory · harness-disclosure · model-or-harness-taxonomy · harness-bench · ddr-bench · scaffold-effects-gaia · awesome-agent-evolution · beyond-pass-at-1 · longhorizon-harness · synthesis