Recoleta
GitHub
Source note

Context Sculpting

Authors
perceptronblues
Source
Hacker News
Published
Collection
Software Intelligence
Agent HarnessesContext ManagementSoftware AgentsCode IntelligenceMulti Agent Systems
Open originalMarkdown source

Summary

Context Sculpting tests a two-model agent harness where a stronger outer model can rewrite a weaker inner model's context between turns. The case study shows that active context rewriting works, but the runs also show high cost, latency, and control-policy risk.

Problem

  • Most chat and agent systems treat the context window as an append-only log. That can preserve stale files, failed paths, tool noise, and distractor evidence, which can waste tokens and steer long-running software agents poorly.
  • The concrete question is whether one model can manage another model's working context by removing misleading state, compacting history, or steering the next step.
  • This matters for code repair and corpus-grounded answers because the model's current context affects what it reads, edits, cites, verifies, and stops doing.

Approach

  • The prototype uses a Pi agent harness with an inner task agent and an outer control agent. The reported runs use gpt-5.4-mini as the inner model and gpt-5.4 as the outer model.
  • After each inner-agent turn, the outer agent reads the full inner context and chooses one action: pass_through, rewrite_context, rollback, or terminate.
  • Demo 1 used two task types: a small file-backed task-manager CLI repair task and a local-corpus synthesis task with distractor documents. It ran 2 inner-only and 2 outer-harness runs per task, for 8 runs total.
  • Demo 2 changed the outer prompt to an intervention_targeted profile, added noisier tasks, and weakened the inner model setting by using low thinking effort.

Results

  • Demo 1 completed 8 of 8 runs successfully, passed 8 of 8 verification checks, triggered 0 guardrails, and cost an estimated $0.7079.
  • Demo 1 did not produce context rewriting. Across 4 harnessed runs, the outer agent made 16 calls: 12 pass_through, 4 terminate, 0 rewrite_context, and 0 rollback. The full harness cost 14x more than inner-only runs without improving pass rate.
  • Demo 2 cost about $1.26 for 4 runs. Across 2 harnessed runs, the outer agent made 15 calls: 14 rewrite_context, 1 terminate, and 0 rollback, which demonstrates active context rewriting.
  • In the noisy synthesis task, the outer agent used an early inject rewrite, later used a compact rewrite, and the inner agent wrote the final answer on the next turn. The excerpt does not report an accuracy delta against a baseline for this task.
  • In the coding repair task, the control run passed in 7 turns, 42.7 seconds, and about $0.015. The harnessed run also passed, but used 12 turns, 566.9 seconds, about $1.06, 12 outer calls, and 12 rewrites, then hit the maxInnerTurns guardrail.
  • The strongest claim is feasibility, not a benchmark breakthrough. The case study shows that prompt policy can change the outer agent from passive supervision to aggressive context editing, and that aggressive editing can make a successful run much more expensive.

Link