Rebuild the workspace, not the agent: companies should store operational context as an LLM-native prose substrate so AI agents read, learn and are governed from the same versioned files; this 'substrate inversion' aims to make deployments auditable, coordinated and capable of compounding improvements rather than stalling after pilot runs.

Main Finding

The paper argues that many enterprise agent failures (governance gaps, poor auditability, coordination drift, safety issues, and stalled improvement) share one architectural root: agents reason over substrates built for humans and traditional apps (fielded schemas, APIs) rather than over representations matched to LLM reasoning. The proposed remedy—"substrate inversion"—rebuilds the cognitive substrate as an agent-native, prose-like corpus (Markdown as an immediate instantiation), isolates schema translation to a single action boundary (the Sync Agent), and places governance, audit, and improvement loops in-band on the same substrate. Two mechanisms (context-bandwidth asymmetry and cross-loop coupling) justify the claim and yield falsifiable hypotheses predicting better compositional reach and sustained compounding improvement when a single, LLM-native substrate is used.

Key Points

• Problem diagnosis

  • Many pilots stall after initial demonstrations; deployed agents rarely improve over time or retain feedback.
  • Common failure symptoms: governance, transparency, coordination, safety, and plateaued improvement—treated as separate but sharing a structural cause.

• Core claim: substrate inversion

  • Choose a single reasoning surface matched to LLMs (connected prose / compiled Markdown) and make schema translation only at the action boundary.
  • Markdown is presented as a practical instantiation today, not a definitive, proven primitive.

• Two mechanisms

  • Context-bandwidth asymmetry: LLMs read long connected prose in one pass, preserving relations implicitly; typed schemas/field-by-field access strip relations and force the agent to reassemble context at every step. Hypothesis H1: per-step reassembly cost bounds compositional reach; the bound tightens with task depth.
  • Cross-loop coupling: sustained improvement needs three coupled loops (action — seconds, skill — days, policy — weeks). When these loops share one substrate, corrections compound; when they are split across different stores, improvements decay. Hypothesis H2: split substrates degrade compound improvement toward noise floor, not rescued by model upgrades alone.

• Operational framework (four layers)

  • Data Layer: authoritative systems of record (CRM, ledger, etc.) left intact—transactional guarantees remain separate.
  • Knowledge Layer: compiled, connected prose (single Git/Markdown repo + knowledge graph) as the shared working memory humans and agents read/write.
  • Intelligence Layer: agents (department, quality, orchestrator, Sync Agent) that coordinate by editing the Knowledge Layer; no out-of-band message bus.
  • Governance Layer: policies, approvals, trust gradients stored in-band in the same Markdown; commits are signed and revertable—policy and auditability become structural.

• Sync Agent and action boundary

  • The Sync Agent is the only component allowed to call external systems (sensor/actuator), performing compilation into the Knowledge Layer and typed writes back to authoritative systems.
  • A per-skill trust gradient and commit-gating rules make trust and approval explicit and versioned.

• Risks and objections

  • New attack surface: indirect prompt injection via compiled external content; the compile path must be secured and audited.
  • Engineering caveats: freshness, read-level authorization, write concurrency, latency/cost of rereading large corpora, recovery from substrate corruption.
  • The paper does not claim empirical superiority; it lays out falsifiable experiments and a benchmark comparing substrates over repeated operation.

Data & Methods

• Type of work: position paper / architectural thesis (accepted to AGENTICS 2026 and on arXiv). No experimental results reported.
• Methods:
  • Conceptual analysis grounded in prior literatures: classical multi-agent systems, blackboard/tuple-space coordination, distributed cognition, and software-engineering technical-debt analyses.
  • Survey/contrast of three current enterprise-agent families: retrieval-augmented agents (RAG), typed-tool agents, and vertical copilots; tabled differences in where reasoning and governance live.
  • Two mechanism-driven hypotheses (H1 and H2) derived from information- and control-theoretic intuitions.
  • A four-layer architectural specification with an illustrative one-page customer file example and intended operational rules (Sync Agent, trust gradient, Git-based audit trail).
  • Critical analysis of objections and risks (prompt injection, governance limits).
  • Proposal of an experiment/benchmark design to evaluate substrates (compare prose-compiled substrate vs retrieval/typed-tool stacks over multi-step, repeated tasks measuring compositional reach, compounding improvement, auditability and safety metrics).
• No empirical datasets used; the paper proposes experiments and metrics for future empirical validation.

Implications for AI Economics

• Deployment economics and ROI

  • Potential to raise the fraction of pilots that reach production by structurally addressing governance, transparency, and compounding improvement.
  • Embedding policy and audit into the operational substrate reduces out-of-band governance costs and lowers coordination overhead between teams (less glue code and fewer one-off translation pipelines).
  • Initial build cost (migrating to a compiled Knowledge Layer, designing Sync Agent, instrumenting commit gates and trust gradients) may be nontrivial; organizations must trade up-front engineering effort for lower long-run operating and compliance costs.

• Technical debt and productivity

  • Substrate inversion reframes "hidden technical debt" (glue code, fragile pipelines) by making the knowledge artifact the canonical asset—reducing repeated translation costs and potential pipeline jungles.
  • If hypotheses hold, agents can achieve deeper, more reliable multi-step behaviors and compounding improvements, increasing sustained productivity gains and reducing labor costs for interpretation-heavy workflows.

• Vendor and market effects

  • Creates demand for agent-native infrastructure: compiled knowledge repos, Sync Agent products, secure compile pipelines, and governance-as-artifact tools.
  • May shift vendor value from isolated tool-APIs toward platforms that enable a unified substrate and integrated governance. Alternatively, raises lock-in considerations around a company’s compiled corpus and toolchain.

• Risk, regulation, and insurance

  • Improved in-band auditability (versioned commits, signed changes) lowers regulatory and legal risk and may reduce compliance costs and liability insurance premiums.
  • New security costs arise (mitigating indirect prompt injection, securing compile paths), which affect operating expenses and risk management pricing.

• Measurement and contracting

  • Encourages substrate-level KPIs (compounding improvement rate, mean-time-to-policy-update, audit reconstruction time) that can be used in SLAs and vendor contracts instead of only per-call accuracy metrics.
  • Suggests an economic case for paying for long-horizon evaluation (multi-step, repeated operation) rather than only single-run benchmarks.

• Scaling and limits

  • Context-window limits imply practical envelopes where single-pass reading is plausible; beyond that, index-first navigation over the same prose substrate is the natural extension. Cost/latency trade-offs and storage/compute for rereading must be considered.
  • For very large companies or very high-volume datasets, hybrid architectures or paging strategies will be needed—so substrate inversion is not a universal panacea but a structural choice with regimes of advantage.

• Practical recommendations for firms and researchers

  • Firms: pilot substrate-inversion on interpretation-heavy, mid-sized operational domains; migrate a bounded set of customer/playbook pages into a versioned prose repo and run A/B tests against retrieval/typed-tool baselines; instrument compounding-of-improvement metrics.
  • Researchers: implement the paper’s benchmark comparing compiled-prose vs retrieval vs typed-tool agents over repeated multi-step tasks; test hypotheses H1 and H2; measure attack surface, governance effectiveness, and compounding rates.
  • Security and governance teams should prioritize hardening the compile path and formalizing trust gradients and commit-gates before large-scale adoption.

Overall, the paper reframes an economic trade-off: invest in a unified, agent-native substrate and pay translation once at the boundary (plus governance and security overhead) in exchange for potentially sustained, compounding agent performance, better in-band governance, and reduced operational technical debt—subject to empirical validation via the proposed benchmarks and experiments.