Business model, goals, roadmap and KPIs in one governed ledger, with a specialist agent for every role, working hand in hand with every employee and every customer.
For years I have worked with some of our largest global technology customers in healthcare and manufacturing. Right now I am helping them use agentic engineering to fundamentally transform how their IT infrastructure is built and run. But I have come to believe that this is the smaller story. What software development is demonstrating today can become the change agent for the entire company. Not by handing every department a chatbot, but through concrete, decisive steps and some structured thinking: setting up an agent harness for the enterprise itself.
A disclosure before I go further: I work at Salesforce, which markets the term „Agentic Enterprise“ [1]. What follows is my personal view, and I keep it deliberately vendor-neutral, built on open standards that no single company owns.
Why agents work in software, and almost nowhere else
At Anthropic, Claude now writes about 80% of the code that gets merged, and engineers ship eight times as much per quarter as they did in 2021 to 2025 [2]. The models got good, yes. But that is not the whole reason it works.
Coding agents became useful because software already had four properties that most of the business lacks:
- One versioned source of truth. The repository.
- Machine-readable intent. Specs, tests and instruction files that say what „done“ means.
- Cheap, fast verification. A CI pipeline that tells the agent within minutes whether it was right.
- An accountable merge gate. A pull request, and a named human who approves it.
Anthropic’s own engineering guidance says agents need „ground truth from the environment at each step“ [3]. In a typical company, the strategy lives in a slide deck from March. The marketing plan lives in someone’s inbox. Three dashboards define the same KPI three different ways. There is no ground truth for an agent to check its work against.
The Agentic Enterprise is not a company with more agents. It is a company that has built these four properties for the business, and only then lets agents work on it.
Company as Code: one ledger for the company’s intent
Company as Code means treating the company’s intent the way engineers already treat their infrastructure: written down, versioned, reviewed and tested. Picture one Git repository. Call it the ledger. It holds the company’s intent, not its transactions:
company/
├── AGENTS.md rules of engagement for every agent
├── constitution/ mission, business model, risk appetite, decision rights
├── strategy/ goals/ corporate strategy, OKRs, strategic bets with kill criteria
├── okrs-kpis/ KPI catalog: definition, owner, target, data source
├── roadmap/ initiatives, each with requirements, design, tasks
├── functions/ marketing, sales, IT, finance … strategies and playbooks
├── policies/ human-readable rules plus their policy-as-code twins
├── agents/ skills/ agent registry, permissions, reusable procedures
└── decisions/ one record per merged decision
What does not belong in it: customer records, invoices, HR files. They stay in their systems of record, and agents reach them at runtime through connectors such as the Model Context Protocol, which respect existing permissions [4]. Git history is effectively permanent, and personal data has a right to be forgotten.
Every file is short, has an owner and a review date. Every change is a pull request. Merged means in force. GitLab has run „all work, process, and policies“ through a merge-request-driven handbook for years [5].
The idea is in the air, and I don’t claim to have coined the term. Daniel Rothmann argued for „Company as Code“ in 2025 [6], Pirol Labs has published an open Company-as-Code standard [7], and Kortix, a vendor, ships an open-source CompanyOS built on the same thesis [8]. Almost all of this work targets startups and small teams. What is missing is the enterprise version: a merge gate for business decisions, a rights matrix, graded human oversight, and answers for the EU AI Act and the works council.
The harness: how agents work on the ledger
A harness runs an agent in a loop with tools, context and permissions. Claude Code and Codex are harnesses for developers; since January 2026, Anthropic has offered the same harness to everyone else as Cowork [9]. The building blocks are open: AGENTS.md and the Model Context Protocol sit under the Linux Foundation [10], and Agent Skills package company procedures [11].
Agents take on familiar roles: an orchestrator turns an OKR into tasks, a critic checks every proposal against strategy and brand, a KPI analyst flags drift.
The central idea: a decision is a merged pull request. It names the goal it serves, the options, the expected KPI impact, the risk and the kill criteria. Automated checks run before any human looks, from policy-as-code [12] to a check that flags work no goal asked for.
Strong platforms, not vibe-coded software
One misreading to rule out: this is not a call to vibe-code the company, prompting an agent until something seems to work. That produces software nobody specified, reviewed or can maintain, and in a regulated business mainly risk.
The companies that succeed with agents do the opposite. They run on strong platforms for everything that must not break: data, identity, security, audit and scale. On top of that foundation, they use agentic coding to build and change their business processes. That is fast, because the platform already brings the data model, access rules and audit trail. And it is safe, because every change still passes a spec, automated checks and a human merge gate.
Speed comes from the platform and the rails, not from skipping them.
Agentic Governance: agents propose, humans dispose
Governance is where agentic programmes quietly fail, so it belongs in the repository, not on top of it.
Rights follow document classes. The board approves changes to the constitution, with four eyes. Executives approve strategy and goals. Function leads approve their own playbooks. Teams approve their operational artifacts. A CODEOWNERS file turns this into required reviewers, path by path [13].
Agents read and propose. They never approve and never merge. GitHub already applies this rule to its coding agent, which „cannot approve or merge a pull request“ [14]. Apply it to the whole company. In the diagram above, no approval line starts at an agent.
No self-modification. No agent may change its own permissions, its own instructions, the policies that check it, or the constitution. Instruction files are an attack surface: researchers have shown hidden Unicode in agent rule files making coding agents insert malicious code [15], and OWASP ranks goal hijack and context poisoning among the top risks for agentic applications [16]. In Company as Code, every AGENTS.md is executable policy.
Break the lethal trifecta. No single agent should combine private data, untrusted content and the ability to communicate externally [17]. Split the reader from the actor.
Autonomy is earned. Each agent runs at a defined human-in-the-loop level, from H0 (drafts only) to H4 (autonomous inside hard bounds). It moves up only after its evals pass and its human owner signs off. One incident moves it down automatically.
The factory: from goal to measured result
With ledger, harness and governance in place, work flows like a production line. Intent, usually a goal or a KPI drift, becomes a spec. The spec becomes a plan with a risk tier. Worker agents execute in sandboxes. Critic and guardrail agents verify, with evidence. A human code owner reviews and merges. The change goes live, the KPI analyst measures it, and what was learned flows back into skills, instructions and evals.
Two feedback loops keep it honest: a fast weekly loop for operational corrections, and a slow quarterly loop in which OKR results become the next strategy release. Gartner found that only 48% of digital initiatives meet or exceed their business-outcome targets [18]. A ledger makes the line from strategy to outcome traceable.
Expect the bottleneck to move to review. On teams with high AI adoption, developers merge 98% more pull requests, while review time rises 91% [19]. McKinsey argues that two to five people can already supervise 50 to 100 specialised agents running an end-to-end process [20], but only if verification is industrialised: risk tiers, findings that carry their own proof, sampled human review. Invest in verification before you invest in generation.
Inside the factory: one request, end to end
The cycle above is the view from altitude. Here is a single request on the line. It is the pattern I see working in software delivery today, stripped of its tooling.
Work arrives through six channels: ideas and the backlog, incident and error reports, routine checks and audits, customer cases and feedback, changes announced by suppliers and platforms, and data from live operations. Intake and triage classify each signal, estimate its size and route it.
A planning agent picks it up, prepared by the ledger and by the company’s standards and rules. Its one hard rule: it briefs specialist agents in writing, through files, and never does the work itself. The specialists run in a fixed order. One researches within the guardrails, one drafts a solution and records its decisions, one builds it from proven patterns, one tests and validates, and one reviews for security, quality and standards. The result is a pull request. If a step fails, the work goes back to the planner, at most twice. After that, a human decides. Iteration is bounded by design.
Four human gates sit on the line:
- G1 Specification: is this the right problem?
- G2 Design: is this the right solution?
- G3 Merge: is the change good enough to accept?
- G4 Release: is it safe to go live now?
After the merge, delivery runs only through the pipeline: version history and automated checks, then test environments, then the release gate, then live. Nobody, not even an agent, changes live directly. Metrics, traceability and evals from operations generate new signals that flow back into intake and close the loop. People and partner teams who take on delegated work ride the same rails, through the same gates and checks. Underneath sits the foundation: skills first, centrally managed settings and enforced checks, a separate identity for every agent, and a complete audit trail.
The pattern comes from software, but nothing in it is specific to code. A marketing brief, a contract amendment or a process update can pass the same four gates. You do not have to trust the individual agent. You trust the rails.
A specialist agent for every role, next to every person
Let me be explicit about where this leads. The Agentic Enterprise does not replace its people with a pool of generic agents. It gives every role in the company its own specialised agents, working hand in hand with the people in that role. Sales gets agents that know the accounts, the pricing rules and the playbook. Finance gets agents that know the closing calendar and the controls. Service, procurement, HR, marketing and operations each get agents built on their own skills, policies and KPIs from the ledger. The employee sets the direction, makes the judgement calls and owns the result. The agent prepares, executes and checks, around the clock.
The same holds outside the company. Customers will work with agents, too: one that resolves a case at night, one that guides a new customer through onboarding, one that knows their contract and history. Because all of them draw on the same ledger, a customer gets the same answer an employee would give. And because they talk to the outside world, they get the tightest bounds: access only to that customer’s data, no write access to the ledger, and a clean handover to a human whenever a case leaves their mandate.
The goal is not a company with fewer people. It is a company in which every employee, and every customer, has a team of specialists at their side.
Change management: the hard part is not technical
DORA’s 2025 research describes AI as an amplifier of an organisation’s existing strengths and weaknesses [21]. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027 [22]. The lesson: the operating model decides, not the model.
I would run the change in three phases:
- Days 0–90: prove one value stream. An executive sponsor and a small ledger council: strategy, platform, compliance and data protection, the works council, two function leads. Write constitution v1, import the current strategy and KPI catalog, switch on CODEOWNERS. Agents may only read and propose.
- Months 3–6: make the loops real. Three to five functions in the ledger, each with its own AGENTS.md and owner. The KPI loop runs weekly; quarterly planning runs as pull requests. The first agents move to H2 after passing their evals.
- Months 6–12: the company is the repo. Everyone can propose. Strategy releases are tagged. Parallel slide decks and wikis are retired.
Three things make or break it. First, leaders write their strategy as pull requests. If the CEO’s strategy lives in a deck, the ledger is dead on arrival. Second, new roles get real mandates: an accountable human owner for every agent, a context engineer who curates the ledger, policy and eval engineers, and a platform team that runs the harness as an internal product. Third, nobody has to learn Git. A web editor or a Cowork-style front-end turns plain language into a pull request.
In Europe, add regulation to phase one. AI literacy under Article 4 of the AI Act already applies [23]. The high-risk obligations were postponed by the Digital Omnibus to 2 December 2027 for Annex III uses such as HR and credit, and to 2 August 2028 for AI embedded in regulated products [24]. In Germany, agent logs that can reveal employee performance require works-council co-determination [25], so negotiate a works agreement early. In healthcare and manufacturing, clinical and safety functions stay inside their certified processes. The ledger runs the operations around them.
An honest caveat
No company I know of runs entirely as a repository yet. This is a proposal built from proven parts, not a reported case study. The productivity evidence is mixed, too: a randomised study by METR found experienced developers 19% slower with AI tools in early 2025, while they believed they were faster [26]. Still, the ledger pays off even without a single agent, because it forces a company to write down what it wants, who decides and how it will know.
Start with the repository
The question for leadership is not which agent platform to buy. It is whether your company’s intent is written down clearly enough that a machine, or a new hire on day one, could act on it. Build the repository. Define the merge gate. Then let the agents in: into every role, next to every employee, and in front of every customer.
Sources
- Salesforce. „What Is the Agentic Enterprise?“ 2026. salesforce.com/agentforce/agentic-enterprise
- Anthropic. „How Anthropic secures its AI-native software development lifecycle.“ 13 Nov 2026. claude.com/blog/how-anthropic-secures-its-ai-native-software-developme…
- Anthropic Engineering. „Building effective agents.“ anthropic.com/engineering/building-effective-agents
- Model Context Protocol. Specification (2025-06-18). modelcontextprotocol.io/specification/2025-06-18
- GitLab Handbook. „Handbook-first.“ handbook.gitlab.com/handbook/company/culture/all-remote/handbook-first
- Daniel Rothmann (42futures). „Company as Code.“ 25 Feb 2025. blog.42futures.com/p/company-as-code
- Pirol Labs. „Company as Code“ — open standard (GitHub). github.com/Pirol-ai/company-as-code
- Kortix. CompanyOS — open-source repository (GitHub). github.com/kortix-ai/CompanyOS
- Anthropic. „Introducing Claude Cowork“ (deployment guide). claude.com/resources/guides/deploying-claude-across-your-organization/…
- Linux Foundation. „Formation of the Agentic AI Foundation (AAIF).“ 9 Dec 2025. linuxfoundation.org/press/linux-foundation-announces-the-formation-of-…
- Agent Skills. Specification. agentskills.io/specification
- Open Policy Agent. Documentation. openpolicyagent.org/docs/latest
- GitHub Docs. „About code owners.“ docs.github.com/en/repositories/managing-your-repositorys-settings-and…
- GitHub Docs. „Risks and mitigations for Copilot cloud agent.“ docs.github.com/en/copilot/concepts/agents/coding-agent/risks-and-miti…
- Pillar Security. „Rules File Backdoor.“ 18 Mar 2025. pillar.security/blog/new-vulnerability-in-github-copilot-and-cursor-ho…
- OWASP GenAI Security Project. „Top 10 for Agentic Applications for 2026.“ 9 Dec 2025. genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-202…
- Simon Willison. „The lethal trifecta for AI agents.“ 16 Jun 2025. simonwillison.net/2025/Jun/16/the-lethal-trifecta
- Gartner. „Only 48% of Digital Initiatives Meet or Exceed Their Business Outcome Targets.“ 22 Oct 2024. gartner.com/en/newsroom/press-releases/2024-10-22-gartner-survey-revea…
- Faros AI. „The AI Productivity Paradox.“ 2025. faros.ai/blog/ai-software-engineering
- McKinsey. „The agentic organization: Contours of the next paradigm for the AI era.“ 26 Sep 2025. mckinsey.com/capabilities/people-and-organization/our-insights/the-age…
- Google Cloud DORA. „State of AI-assisted Software Development 2025.“ dora.dev/research/2025/dora-report
- Gartner. „Over 40% of Agentic AI Projects Will Be Canceled by End of 2027.“ 25 Jun 2025. gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-ove…
- EU AI Act, Article 4 (AI literacy). artificialintelligenceact.eu/article/4
- European Parliament, Legislative Train. „Digital Omnibus on AI“ (Reg. (EU) 2026/1744). europarl.europa.eu/legislative-train/package-digital-package/file-digi…
- Betriebsverfassungsgesetz, §87 (co-determination). gesetze-im-internet.de/betrvg/__87.html
- METR. „Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity.“ 10 Jul 2025. metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study
All links checked on 17 November 2026. Gartner and McKinsey block automated checks (HTTP 403 / timeout); both pages were confirmed through search.


