Agent Substrate

Reference entry · last updated September 3, 2026

Agent substrate is the part of agent infrastructure that describes the system or domain: its data model, specifications, decisions, conventions, and private context. It changes when the work changes.[1]

In Nestor G Pestelos Jr's implementation, the substrate is durable, operator-owned, and model-independent. Multiple vendor runtimes can use it without renegotiating the system's rules. This extension describes the author's framework and is not established industry terminology.

Components

A substrate records facts about the work and the rules that govern it. It can contain data models, specifications, decisions, conventions, and private context.[1] In this framework it can also include safety constraints and operational procedures. This information remains useful when the model or vendor runtime changes.

The extension to safety constraints and operational procedures describes the author's implementation.

Relation to an agent harness

An agent harness is a runtime such as Claude Code or Codex. It supplies the model, tool interface, execution loop, and session controls. The substrate is the owned layer that these harnesses read, write, and operate on.

Several harnesses can use the same domain model and operating rules. Replacing a harness does not require the operator to redefine the system.

This boundary is part of Pestelos's framework, not a general definition accepted across agent platforms.

Substrate and scaffolding

Substrate changes when the represented work changes. Scaffolding compensates for a limitation in a model or runtime, so a model release can make it obsolete.[1]

Anthropic described this effect in a long-running application-development harness. Each harness component encoded an assumption about model limits. The team removed its sprint construct for Claude Opus 4.6 and moved the evaluator to one final pass. It retained the planner. The builder then ran for more than two hours without sprint decomposition.[2]

Chroma reported that model performance often declines as input length grows, although the shape and degree vary by model and task.[3] A context rule belongs to the substrate when it expresses a durable domain requirement. It is scaffolding when it only compensates for current model behavior.

Mixed artifacts

One file can contain both layers. A memory file may hold a durable project decision beside a temporary compaction rule. A pipeline may encode an operator's approval boundary and also divide work into smaller steps because a current model loses track of long tasks.

The classification follows purpose rather than file type. Separating durable information from temporary compensation makes obsolete scaffolding easier to remove.

Examples

vault conventions that define names, relationships, and storage rules;
a Workstreams ledger that records commitments and their state;
Git guards that preserve operator-defined safety boundaries;
Git-tracked memory shared across sessions;
skills that encode repeatable procedures; and
pipelines and loops that preserve durable operating rules.

These examples describe the author's implementation of the framework.

Limits

Durability tracks stable domain rules. The substrate must be updated when those rules change.
The format must remain compatible with every harness that reads or updates it.
Ownership requires control over meaning, access, and migration. Storage can be local or remote.
Correctness requires maintenance and validation. Durable storage preserves stale facts as stale facts.

References

↑ Nestor G Pestelos Jr, “Your Agent Infrastructure Has Two Lifespans,” revised September 3, 2026. https://ngpestelos.com/writing/half-agent-infrastructure-depreciating-asset/
↑ Prithvi Rajasekaran, “Harness design for long-running application development,” Anthropic Engineering, March 24, 2026. https://www.anthropic.com/engineering/harness-design-long-running-apps
↑ Kelly Hong, Anton Troynikov, and Jeff Huber, “Context Rot: How Increasing Input Tokens Impacts LLM Performance,” July 14, 2025. https://www.trychroma.com/research/context-rot