Spine: Compiling Subtask Trees into Structured Context for Long-Horizon Coding Agents
Long-horizon coding agents bloat context, and trimming loses unfinished work. Spine compiles a subtask tree per call: +11.8% score, 32.4% less cost.
The claim
Spine compiles a subtask tree into context, so coding agents score higher at lower cost.
The problem
Long-horizon coding tasks contain many nested subtasks that build on each other.
Sequential history keeps growing until context inflates and model quality degrades.
Trimming that history can discard information unfinished subtasks still need.
How it works
Spine lets the agent declare explicit subtask boundaries through work scopes.
One agent runs every nested subtask inline instead of delegating each one.
Delegation stays reserved for independent tasks that can run in parallel.
A persistent SpineTree records subtask relations, interactions, completion states, and returned results.
Before each model call, the runtime compiles the tree just in time.
Figures
What changes
Completed subtasks reduce to their results while unfinished ones keep full detail.
SpineCodex lifts SWE-Milestone Score by 11.8% over native Codex.
Total cost falls by 32.4% while ProgramBench test-pass rate rises 17.3%.
Limits
Most results come from benchmarks rather than live repositories.
The agent must declare work scopes for the tree to help at all.
Headline results use GPT-5.6 Sol, with two open models following on Pi.
Source · arXiv