Nestor G Pestelos Jr · Writing · Print
Your Agent Infrastructure Has Two Lifespans
Agent infrastructure has two clocks. Substrate encodes your problem and changes when the problem changes. Scaffolding compensates for a model limitation and can expire with the next release.
The September 2, 2026 version is at half-agent-infrastructure-depreciating-asset-20260902.
Most advice about AI agents recommends better instructions and structured context inside a harness. That advice mixes two kinds of infrastructure with different lifespans.
Substrate describes your problem: the data model, specifications, decisions, conventions, and private context. It changes when your work changes.
Scaffolding compensates for a model limitation. It includes task decomposition, tool routing, context-reset rules, and other controls that keep a model on track. A new model can make those controls unnecessary.
Two clocks
Anthropic documented the difference in a long-running app harness. With Opus 4.5, the team split work into sprints and graded each sprint. After moving to Opus 4.6, it removed the sprint construct and moved the evaluator to one final pass. It kept a planner because the raw prompt led the generator to under-scope the app. The builder then worked coherently for more than two hours without sprint decomposition.
That makes scaffolding a depreciating asset. You may need it now, but its useful life can depend on a vendor's release schedule. This is one documented case, not a measured decay rate.
Better models still need relevant context because they do not know your private decisions or repair an incoherent specification. Chroma's report across 18 language models found that performance became less reliable as input length grew, even on simple tasks, so organized context still matters.
I argued in Cheap Code, Expensive Context that value moves to whoever holds the context.
One file, two lifespans
My agent memory contains both kinds. The facts record what is true about my system, which conventions apply, and why I made past decisions. Those facts have survived model upgrades this year.
The rules for compacting that memory and recovering after a context reset are scaffolding. They exist because current models lose or misuse context. I expect to delete some of those rules as the models improve.
A prompt, script, or database can contain durable knowledge and temporary compensation at the same time.
Whether agents eventually absorb the substrate work too is still open. Until then, the operator owns failures caused by stale context, hidden decisions, or conflicting instructions.
A deletion test
Before adding structure, ask what it encodes.
- If it describes your domain, invest in making it clear and correct.
- If it compensates for a model failure, record the assumption behind it and keep the component easy to remove.
- If it does both, separate the durable information from the temporary control logic.
After model upgrades, retest the temporary parts and remove any that no longer improve results.
I spend design effort on the substrate and keep the scaffolding disposable.
Sources
- Prithvi Rajasekaran, Harness design for long-running application development, Anthropic Engineering, March 24, 2026
- Kelly Hong, Anton Troynikov, and Jeff Huber, Context Rot: How Increasing Input Tokens Impacts LLM Performance, Chroma Research, July 2025