How to run a company as a MAPE-K loop

How to run a company as a MAPE-K loop

To run a company as a MAPE-K loop, you give every recurring business process four separable jobs (watch it, judge it, decide a change, apply the change) and one shared, durable record that all four read and write. The loop is the firm's operating unit; the people or agents inside it are the sensors, analysts, planners and effectors. The 561 Group has recorded as a standing requirement that it wants "to move more and more toward treating things like a MAPE-K loop," and Coesite Operations, which runs the Group's operating loops, is where that requirement is worked out.

What is MAPE-K?

MAPE-K stands for Monitor, Analyze, Plan, Execute over a shared Knowledge base. It is the canonical reference architecture of autonomic computing, the program that IBM launched in 2001 and that Kephart and Chess described in 2003 (Kephart & Chess 2003)¹¹ Jeffrey O. Kephart and David M. Chess. "The Vision of Autonomic Computing." Computer 36, no. 1 (2003): 41-50.. An autonomic manager sits next to a managed resource. Sensors expose the resource's condition, the monitor turns signals into observations, analysis decides whether adaptation is required, planning builds a change, execution applies it through effectors, and the knowledge base holds the models, policies, history and goals the other four need (IBM Corporation 2005)²² IBM Corporation. An Architectural Blueprint for Autonomic Computing. IBM, 2005..

Lenticule Science's research note on the subject makes three points worth carrying into operations. MAPE-K is not an algorithm: it decides the division of labor, not whether analysis is rules or statistics. The letters need not be five separate services: a realization may fuse them. And the K is not a database bolted on; it is the shared representation that makes the four activities mutually intelligible. The loop can react without it, but cannot reliably regulate. The research itself is at Lenticule Science.

Why would a firm, not just a server, use it?

Because a firm already has the four jobs and usually does them badly. The monitoring is a dashboard nobody reads, the analysis is a meeting, the plan is a ticket, the execution is whoever is free, and the knowledge is a chat history. MAPE-K gives each job an owner and a record, and it names the loop that closes: what was executed changes what will be monitored next.

The deeper reason is cost. A loop whose stages can run without a person, or without a language model, is a loop that can run every hour. The Group's rule is explicit that recurring work moves into deterministic software "so it runs without a model in the loop," with typed operations and retained original evidence, rather than being re-orchestrated by an agent session each time. Model inference is for judgment, which is mostly the Analyze and Plan stages on novel cases. See deterministic tools before model tokens.

How do the five stages map onto a company?

We describe each stage below with its evidence rule.

Monitor: what does the firm see?

Monitoring in a firm is not "the build is green." It is a reading taken from outside the system, with a time and a method. The Group's own estate probe is the clearest example: a Worker GETs each public node without following redirects, times out after four seconds, and tells each answer in Belnap's four values. A node that answers with 401, 402 or 403 is told both: answered, and refused. We explain this in four truth values for a service estate. The point for MAPE-K is that the monitor's output type is richer than up/down, because a binary monitor cannot represent a paywall, a login wall or a monitoring failure honestly.

A monitoring failure is itself an observation. If the probe cannot reach a server, that proves the observation failed; it does not by itself prove the server is down.

Analyze: where did the chain stop?

Analysis in operations asks for the earliest transition whose result we cannot demonstrate. A change can exist in a source revision, be absent from the build artifact, be present in an artifact the provider never accepted, or be accepted by a provider but not routed to the hostname. Each check narrows the intervention. This is the logic of the evidence ladder for a service that reached someone.

Where two offices disagree about reality, the Group's rule sends the disagreement to a formal coordination framework that preserves enough provenance to re-derive the decision, instead of flattening it to one priority number such as jurisdiction or the count of evidence.

Plan: who owns the change?

The plan is a work order, and its form is regulated. A request's order carries what any carrier needs to make it, namely requirements, prohibitions, the result's type and admitted examples, and nothing about who makes it. Before a session changes anything it must finish context discovery (the governing theory, the owning section and office, the existing code and the open work on it), and it fixes its verifier before the first edit. That rule is what keeps Plan from being improvisation.

Decisions about spending, vendors and technical route belong to the office that owns them. They are not passed up to the top of the firm as questions. See agents hold offices.

Execute: what leaves evidence?

Execution is the stage most firms under-record. The Group's decision note on how a node changes itself states the principle: every recurring trigger is declared in the release as a scheduled job, one file per job, with declared working-tree custody so two jobs cannot silently write the same paths. Hand-linked timers, which had accumulated and in several cases ran beside their own replacements, are retired.

Knowledge: what is the K?

Knowledge in the firm is the set of records that a later office can re-read without reconstructing a conversation: incident records, standing rules, glossary definitions, and the law that governs a work arena. Two properties make it fit for a loop. Records are content-addressed (named by a hash of their bytes), so a name cannot be quietly re-pointed. And law is versioned as editions, each naming its predecessor, with each stored move read under the edition that admitted it. How a node changes itself covers this.

What is the difference between a loop and a bureaucracy?

A bureaucracy has stages; a loop has closure. The test is whether the output of Execute becomes an input to Monitor. If a deployed fix is never re-observed from outside, the firm has a pipeline, not a loop. The Group's incident procedure requires that after an intervention the same external test be repeated and its response evidence preserved, appended to the same accountable incident rather than summarized anew by four independent writers.

Where does a loop fail?

Lenticule Science's synthesis names the central difficulty: the boxes are the easy part. The decisive problem is proving properties of the coupled system formed by the boxes, the managed process, the environment and the knowledge they jointly build, including drift, contradiction and path dependence in the K. We do not claim to have proved such properties for the firm. We claim a discipline: separate the jobs, type the observations, retain the evidence, and make correction cheap.

Specific failure modes we guard against:

1. Metric capture. Monitoring what is easy (build status) instead of what matters (what arrived).
2. A K that only grows. Data no consumer reads is deleted rather than kept; unread knowledge is liability.
3. Self-edit without authority. Ordinary method improvements fall within an office's discretion, but a change to constitutive commitments needs a governing act, not a loop that edits itself.
4. Nested loops fighting. Several controllers with different time horizons can destabilize each other; offices have distinct duties precisely so each loop has one owner.

Is this implemented?

Some of it. The estate probe is described by its own repository documentation as running on request with a five-minute cache. The job-scheduling and law-edition decisions were recorded as a decision note in October 2026 with implementation slices landing. The requirement to treat the firm's loops as MAPE-K is recorded, and the Group's homepage states that the whole design is intent and that whether any part is running is a separate question answered by a dated record. Read this article as an operating method with named parts, not as a claim that every part is live.

Where to go next

MAPE-K, defined · the design sentence, explained · Rendanheyi for AI agents · the 561 Group · Red Cup Engineering, which supplies the software that interprets these methods.