Repo-Native Agent Operations (v1.1.0 — archived)

The Bussetech Software Studio Architecture — one operator, a governed gnome workforce, run entirely through a Git hosting platform. This is versioned content; see the changelog for what moves between releases.

Abstract

We describe the architecture and first operating evidence from the Bussetech Software Studio: a one-operator software studio whose workforce is a fleet of governed AI agents — gnomes — managed entirely through a Git hosting platform. The industry’s agent conversation is dominated by orchestration frameworks and observability SaaS; production outcomes remain poor, with 88% of enterprise agent pilots failing to ship and evaluation, governance, and reliability cited as the leading blockers (Turion, 2026). Our position is that these are not tooling gaps but system-of-record gaps. We present repo-native agent operations: a pattern in which the repository platform is the sole message bus, governance layer, memory substrate, and financial ledger for an agent organization. We detail six architectural commitments — the gnome contract, propose-don’t-decide, the two-layer output contract, trust-tier manifests, runner-level cost receipts, and the guidance ledger — and evaluate them against a live deployment that bootstrapped itself in thirteen agent sessions and shipped a self-maintaining data product in the same week. We situate each claim against prior art and are explicit about which elements are novel synthesis versus established practice.

Provenance. Every operational number in this paper traces to the studio’s own repositories, decision records, and run ledger. The companion research notes (market context, the solo-studio model, client positioning) and the positioning theses are internal to the studio.

1. Introduction

Two conversations dominate contemporary software-architecture discourse around AI. The first is capability: coding agents now open credible pull requests. The second is operability: how organizations run such agents continuously without losing control of quality, cost, or intent. The capability conversation has largely been won; the operability conversation has not. Gartner expects more than 40% of agentic AI projects to be canceled by 2027, and enterprises report evaluation gaps (64%), governance friction (57%), and reliability concerns (51%) as their principal blockers (Digital Applied, 2026; Turion, 2026).

The studio’s founding observation is that most agent deployments bolt governance onto an execution stack as an afterthought — a tracing SaaS here, a human-approval Slack message there — leaving no single place where an auditor, a client, or the operator long after the fact can reconstruct what happened and why. We inverted the priority: choose the system of record first, and make every other component subordinate to it. Our system of record is the Git hosting platform (GitHub, in our deployment). Not merely for code, and not merely for configuration as in GitOps, but for all of it: work assignment, agent output, human decisions, escalations, distilled policy, simulation results, and per-run economics.

A second, smaller inversion concerns vocabulary. The industry cannot agree on what an “AI agent” is, and the term’s ambiguity leaks into architecture reviews and client conversations alike. We sidestepped the definitional debate by adopting our own term with a precise local meaning: a gnome is a single agent that handles one task end to end, defined by a manifest contract rather than by anyone’s taxonomy.

2. Related work and prior art

We build deliberately on established practice and claim novelty narrowly.

Agent orchestration. Frameworks such as LangGraph, CrewAI, and AutoGen coordinate multi-agent workflows in application code. The studio intentionally has no orchestration framework: coordination happens through repository primitives (issues, PRs, labels, scheduled workflows). This trades expressive orchestration for total auditability — a trade we defend in §7.

Coding agents and PR-based output. Producing agent work as pull requests is established (GitHub Copilot’s coding agent, Cognition’s Devin, SWE-agent lineage). Our contribution is not PR-based output itself but making PR-only output a manifest-declared, structurally enforced trust property (§4.3) rather than a product behavior.

Observability and cost tracking. A rich tool ecosystem (Langfuse, LangSmith, Helicone, Phoenix) instruments traces, tokens, and cost, with the OpenTelemetry GenAI semantic conventions emerging as a standard. These systems assume an external observability plane. We journal at the single execution path and commit receipts to the repository itself (§4.5) — a deliberately smaller, auditable-by-git alternative for organizations whose scale permits it.

Agent memory. The memory literature has converged on working/episodic/semantic/procedural taxonomies, with OS-style context management popularized by MemGPT and productized by Letta, Mem0, LangMem, and Zep. Our guidance ledger (§4.6) is best understood as procedural memory with human provenance: policy is distilled from human rulings, versioned in git, reviewed like code, and injected at run time. It learns nothing on its own — by design — and every entry cites the decision that created it.

Decision records and platform engineering. Architecture Decision Records are two decades of established practice (Nygard, 2011). We extend the genre with two adjacent artifact types: guidance entries (operational rulings binding on agents) and handoffs (session-to-session state transfer documents), and we hold agents to the ADR discipline that human teams routinely aspire to and skip.

3. Architecture overview

The studio comprises repositories under one GitHub organization, separated into a control plane, an execution layer, and a publish plane. A single human — the sysop — reviews, merges, and rules; everything else proposes.

Three properties follow from the topology. First, the portal is a renderer, not a source: www derives everything from control-plane state at build time, so adding a project requires zero portal edits. Second, portability is contractual: every repository documents a detach procedure and builds without studio access; gnome code contains no hardcoded organization facts (CI greps for violations). Third, visibility is a first-class field: public, private, and private-published (public site, private source) are declared per repository and enforced by leak checks in CI — the mechanism by which public transparency and NDA client work coexist.

4. The six commitments

4.1 The gnome contract

A gnome is a manifest (gnome.yml), a prompt, fixtures, and a thin workflow wrapper. The manifest declares its display name and variable name (gn_<domain>_<action>), level (platform or project), model, token budget, schedule, secrets, deployments, and — centrally — its input-trust tier. Execution is one runner, two runtimes: the same gn-run code path serves scheduled Actions runs and interactive console sessions, which is what makes receipts and policy injection inescapable rather than aspirational.

4.2 Propose, don’t decide

No gnome merges, closes issues, triggers project creation, or posts externally. The management gnomes (backlog refinement, delivery flow, product judgment) draft and escalate; branch protection makes the rule mechanical rather than cultural — during the pilot, even the operator’s console required explicit administrative override to merge, making every merge a deliberate act.

4.3 Trust-tier manifests

input_trust: untrusted is a manifest field with structural consequences: PR-only output, no secrets in context beyond the model key, and fetched content handled as data rather than instructions, with injection lines included in the gnome’s own test fixtures. Prompt-injection posture is thereby declared and enforced outside the prompt. To our knowledge, trust tiering as a manifest-level contract enforced by the execution harness — rather than as per-prompt hygiene — is not described as an integrated pattern in the current agent-framework literature, though it composes ideas well established in security engineering.

4.4 The two-layer output contract

Layer one: the runner validates output shape (declared paths, segment-safe glob patterns with traversal checks). Layer two: the target repository owns semantic validation — JSON-Schema data CI plus a repo-specific referential-integrity hook. The separation matters: the platform cannot know every project’s data semantics, and the project cannot be trusted to re-implement transport safety. The pattern fired in production during the pilot: the scout gnome’s first live output was schema-invalid, data CI blocked the PR with a readable error, the prompt was hardened, and no invalid record ever reached main. We regard this two-layer split — harness shape-check plus consumer-owned semantic CI — as the load-bearing safety property for dataset-producing agents.

4.5 Receipts at the runner

Every run — scheduled or interactive, success or failure, live or dry-run — appends a JSONL receipt (model, tokens, estimated cost, duration, outcome, trigger, runtime) to a ledger branch of the control repo; a scribe gnome aggregates weekly reports and raises threshold alerts (80% of budget; 3× trailing-average spikes). This differs from the observability-SaaS pattern in locus and audience: receipts live in the organization’s own git history, reviewable by the same mechanics as everything else, at zero additional infrastructure. The pilot’s headline economics were legible precisely because of this: a 33-record, 183-signal sourced dataset cost ≈ $4.48 of gnome spend to research, normalize, and publish. We do not claim the approach scales to high-frequency agent fleets; we claim it is correct-sized for studio-scale operations and strictly more auditable.

4.6 The guidance ledger

When a gnome (or session) hits a question it cannot answer, it files a decision issue carrying context, options, a recommendation, and a deadline with a default action — unanswered questions degrade gracefully instead of deadlocking. When the sysop rules, a workflow distills the ruling into a numbered, scoped guidance entry; the runner injects applicable entries into future runs. A question answered once becomes policy.

Five entries existed at pilot close, and the loop ran on live decisions — including a data-modeling ruling (“negative outcomes are records, not omissions”) that now governs every applicable gnome. Read against the memory literature (§2), this is procedural memory that is human-authored, git-versioned, and citation-bearing: slower than learned memory, and deliberately so, because in a client-facing studio the provenance of policy is worth more than the latency of acquiring it. We believe the closed loop — decision issue → ruling → distilled artifact → runtime injection, entirely in repository primitives — is the studio’s strongest candidate for a genuinely novel operational pattern.

5. Organizational constructs

Knolls. Gnomes are grouped into knolls — teams with a shared, knoll-scoped knowledge base built on the same distill-and-inject machinery as guidance. As of this version the knoll framework has shipped (ADR-0033): three knolls are live — a studio management knoll, a kdc project knoll, and the GTM knoll (seven gnomes); the registry’s level field (platform vs project) separates studio-serving from product-serving gnomes, and the guidance injection seam it reuses is live today.

The personality firewall. For brand purposes, gnomes will carry personality overlays derived from their actual run history (a Tamagotchi-inspired score keyed to lifetime tokens). The architectural commitment is a hard firewall: personality artifacts live outside gnome context paths and the harness must never load them — display-only by construction, verified in CI, so relatability can never contaminate behavior. We flag this as a design principle with the build pending.

The sysop console. There is no bespoke dashboard. A documented session pattern (standup / triage / dispatch / inspect) makes any fresh agent session in the control repo a complete operations console, with the rule that every outcome is written back to the repository before the session ends. Terminal scrollback is not a record.

6. Evaluation: the bootstrap and the pilot

The platform was built by thirteen structured agent sessions over three days, each ending in a handoff document; the pilot project (kdc, a US data-center tracker) was then founded by the platform’s own machinery: templated brief → founder gnome’s reuse analysis (which improved on the human brief, folding a proposed third gnome into a mode of an existing one) → repository, DNS, CI, and portal presence with no hand-edits → two project gnomes producing sourced records through the full safety stack. Numbers at pilot close (early July 2026): 6 repos, 11 registered gnomes (100% dry-run simulation coverage), 20 ADRs, 5 guidance entries, 33 site records from 183 per-source signals across 7 states, every fact source-cited, honest negatives included. The receipts have kept accruing since: as of this version (2026-07-06) the studio stands at 7 repos, 18 gnomes, 34 ADRs, 8 guidance entries, and 36 records — the same machinery, more evidence, every count verifiable by git log (platform counts against the platform repository; records/signals against the kdc repository).

The most instructive result was negative: the simulation layer validated the happy path, and all four platform bugs found during the pilot were input-texture failures (punctuation in free text, unquoted dates, token-URL remotes, a vendor CORS outage) that bland fixtures could never reach. The corrective — an adversarial fixture policy enforced in CI — is now standing platform law. We offer this as a general finding for the agent-evaluation discourse: fixture blandness is a coverage gap distinct from scenario coverage, and live use will find it.

7. Limitations

We state these plainly. (1) Scale: one operator, seven repos, eighteen gnomes at the time of writing; none of the receipt or governance claims have been tested at enterprise fleet scale, and the in-repo ledger would need the artifact-collector variant well before high-frequency operation. (2) Platform coupling: the pattern is Git-host-shaped; we use GitHub primitives heavily, and while the portability contract covers project repos, the operational fabric itself would need porting work. (3) No orchestration expressiveness: repository primitives cannot express rich inter-agent protocols; we consider this a feature at studio scale and an open question beyond it. (4) Model dependence: gnomes ride frontier commercial models; routing mitigates cost, not dependence. (5) Novelty: most individual elements have prior art (§2); our claims rest on the integration and on operating evidence, not on invention of parts. (6) Solo-operator bus factor: mitigated by the property that the studio’s entire state is legible from its repositories — the successor’s onboarding document is the organization itself.

8. Future work

Three items on the original future-work list have since shipped and now carry their own operating evidence: autonomous research intake behind an allowlisted, deterministic fetch layer; model routing with fixture-gated downgrades — a downgrade may not regress a gnome’s own fixtures before it takes effect; and the knoll knowledge-base framework (ADR-0033), now live across a studio, a kdc, and a GTM knoll. What remains genuinely ahead: a multi-tenant SaaS stratum to be proven on the studio’s own client portal before any client engagement depends on it; the display-only personality overlay under its CI-enforced firewall; and brownfield adoption — an intake gnome that classifies and onboards existing repositories into the governance fabric.

9. Conclusion

The delivery gap in agentic software is a governance and record-keeping gap. By making the repository platform the message bus, the memory, the ledger, and the constitution of an agent organization — and by holding agents to propose-don’t-decide with structurally enforced trust tiers — a single operator can run a software studio whose every claim is checkable by git log. The studio is small. The pattern is not.

References

Primary external sources: Turion — State of AI Agents in Enterprise 2026 · Digital Applied — AI Agent Adoption 2026 · BCG — The $200B AI Opportunity in Tech Services · AIMultiple — Agent Observability Tools 2026 · SigNoz — LLM Observability Tools · MemGPT, arXiv:2310.08560 · Nygard — Documenting Architecture Decisions · GitOps.tech. Internal sources: the platform repository (ADRs, guidance entries, the EPIC1 and EPIC2 retrospectives, and the run ledger) — the studio’s claim is that these are checkable, so they are the citation.

Versioning & changelog

This paper is versioned content. Semantic versioning applies to its claims: patch = corrections and link fixes; minor = an evidence refresh (live numbers, a new citation or section, a pending item marked shipped); major = an architectural claim changes. This URL (/whitepaper/) always serves the latest version; prior versions are archived at /whitepaper/vX.Y.Z/.