The Agentic Company OS: Substrate Inversion for Sustained Enterprise Agent Deployment
Oliver Aleksander Larsen, Mahyar T. Moghaddam
cs.SE · cs.AI · cs.MA
Quick verdict
Position paper: enterprise agent stalls stem from human-first data substrates; rebuild them as Markdown for LLMs. Lucid, honest, but wholly unvalidated.
Summary
This position paper argues that enterprise AI agents often stall in sustained operation for an architectural rather than a model-capability reason: they reason over working context structured for human operators and conventional applications. The authors propose “substrate inversion” — rebuilding a company’s cognitive substrate as connected prose (Markdown today) that a model reads in one pass, with all schema translation confined to a Sync Agent at the action boundary. Two mechanisms anchor the argument: context-bandwidth asymmetry (typed, field-by-field access strips relations and forces per-step reassembly) and cross-loop coupling (action, skill, and policy improvement loops compound only on a shared substrate). A four-layer framework (Data, Knowledge, Intelligence, Governance), a four-rung per-skill trust gradient, and a constructed customer-file example operationalize the position. The authors state two falsifiable hypotheses and specify a substrate-controlled benchmark, but report no measurements — they concede superiority over retrieval, typed-tool, and hybrid stacks “remains a hypothesis, not a result.” The paper is notably candid: Markdown is “not a proven agent-native primitive,” the example is constructed, and freshness, read-level authorization, concurrency, and compile-path prompt injection remain open. The core limitation is that this is an argument, not evidence.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- The five recurring enterprise-agent failure modes (governance, transparency, coordination, safety, plateaued improvement) share one architectural root: agents reason over working context structured for humans and conventional applications, not for LLMs.
- Rebuilding the cognitive substrate as connected prose (Markdown today) with schema translation isolated to a Sync Agent at the action boundary addresses four of the five failure modes directly and reshapes the fifth.
- Context-bandwidth asymmetry: typed field-by-field access strips relations and forces per-step reassembly, so reassembly cost bounds compositional reach, tightening with task depth (H1).
- Cross-loop coupling: action, skill, and policy improvement loops compound only if they share one substrate; on split substrates improvement decays toward the noise floor and model upgrades alone do not rescue it (H2).
- Governance and auditability can be structural properties of the substrate (version-controlled policy, per-skill trust gradient with shadow/assist/autonomous-sampled/full rungs), though safety still requires enforcement outside the LLM at the action boundary.
Strengths
Genuinely steelmanned alternatives with explicit concessions
Sect. 5 gives each rival position (better retrieval, typed-tool agents, vertical apps) its strongest case and concedes ground precisely: at low compositional depth the approaches are ‘functionally equivalent’ and the choice ‘economic rather than architectural’ (Sect. 5.1); single-vertical companies are ‘well served’ by vertical apps (Sect. 5.3). The hybrid that versions playbooks in one repo is acknowledged to already capture cross-loop coupling (Sect. 5.2).
Falsifiability discipline rare in position papers
H1 and H2 (Sect. 3) come with explicit falsification conditions, and Sect. 7 specifies a 5-representation × 2-coupling factorial benchmark compiled from one source corpus, with compositional depth defined (inter-entity relations jointly resolved, source-fact count and tokens held fixed), a noise floor (variance under a frozen playbook), and a compounding criterion (positive pass-rate trend over tens of closures).
Candid self-identified security analysis
Well-grounded lineage and honest treatment of counterevidence
Weaknesses
Zero empirical content
Motivating evidence rests on one industry report
Mechanism 1’s applicability envelope is undefined
Unsourced quantitative asides
Governance slogan overreaches its own mechanism
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a position paper (accepted at AGENTICS 2026, Springer CCIS) making an architectural, not empirical, contribution. The core claim (Sect. 1): five recurring enterprise-agent failure modes — governance, transparency, coordination, safety, plateaued improvement — are ‘five symptoms with a common architectural root,’ namely that agents reason over working context structured for human operators and conventional applications. The remedy, ‘substrate inversion,’ commits to two structural properties: one agent-native reasoning surface (connected prose; Markdown today, explicitly ‘not a proven agent-native primitive’) and schema translation isolated to a single Sync Agent at the action boundary. Two mechanisms ground the argument (Sect. 3): context-bandwidth asymmetry (H1: per-step reassembly cost bounds compositional reach) and cross-loop coupling (H2: on split substrates, improvement ‘decays toward the noise floor’). Table 1 contrasts the three deployed stack families against the proposal across reasoning surface, translation locus, loop sharing, audit locus, and governance locus. The paper is scrupulous about epistemic status: superiority over typed-tool, retrieval, or hybrid stacks ‘remains a hypothesis, not a result’ (Sect. 8), and ‘Neither outcome is yet measured’ (Sect. 7).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is ‘substrate inversion’ in this paper?
The proposal that companies deploying AI agents rebuild the working context their agents read — currently structured for human operators and traditional applications — into one shared, LLM-friendly corpus of connected prose (Markdown today), keeping all translation to external systems at a single boundary component called the Sync Agent.
Does the paper present experiments or results?
No. It states two falsifiable hypotheses (H1 on compositional reach, H2 on improvement compounding) and specifies a benchmark that would vary representation and loop coupling from one fixed source corpus, but explicitly reports no measurements; the authors concede superiority over retrieval or typed-tool stacks ‘remains a hypothesis, not a result.’
What risks do the authors acknowledge?
Indirect prompt injection on the compile path (persistent poisoning re-read by every agent, cross-agent propagation, loop laundering), an internet-facing unauthenticated email write path, read-level access control in a shared repository, conflict between data-erasure law and immutable Git history, plus open questions on freshness, write concurrency, and cost of rereading the corpus.
Who is this paper relevant for?
Researchers and architects in enterprise agent deployment, business process management, and agent governance. The authors explicitly exclude high-frequency transactional systems (trading, ad bidding, payment routing), where typed guarantees and latency dominate.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–3 appear only as captions and in-text description in the extracted text; the rendered graphics themselves are not part of the text, so figure-level observations rely on the captions.
- The metadata abstract is cut mid-sentence, but the embedded full text contains the complete abstract and the complete reference list (refs 1–47); no truncation affects this review.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.