One of many ideas in our upcoming e book Structure as Code is the flexibility for architects to design automated governance checks for vital architectural considerations, creating quick suggestions loops when issues go awry. This concept isn’t new—Neal and his coauthors Rebecca Parsons and Patrick Kua espoused this concept again in 2017 within the first version of Constructing Evolutionary Architectures, and plenty of of our purchasers adopted these practices with nice success. Nevertheless, our most bold targets have been largely thwarted by a standard downside in fashionable architectures: brittleness. Happily, the arrival of the Mannequin Context Protocol (MCP) and agentic AI have largely solved this downside for enterprise architects.
Health Features
Constructing Evolutionary Architectures defines the idea of an architectural health perform: any mechanism that gives an goal integrity examine for architectural traits. Architects can consider health features kind of like unit exams, however for architectural considerations.
Whereas many health features run like unit exams to check construction (utilizing instruments like ArchUnit, NetArchTest, PyTestArch, arch-go, and so forth), architects can write health features to validate all kinds of vital checks…like duties usually reserved for relational databases.
Health features and referential integrity
Contemplate the structure illustrated in Determine 1.
In Determine 1, the workforce has determined to separate the information into two databases for higher scalability and availability. Nevertheless, the widespread drawback of that strategy lies with the truth that the workforce can now not depend on the database to implement referential integrity. On this state of affairs, every ticket should have a corresponding buyer to mannequin this workflow accurately.
Whereas many groups appear to suppose that referential integrity is just doable inside a relational database, we separate the governance exercise (knowledge integrity) from the implementation (the relational database) and understand we are able to create our personal examine utilizing an architectural health perform, as proven in Determine 2.
In Determine 2, the architect has created a small health perform that screens the queue between buyer and ticket. When the queue depth drops to zero (that means that the system isn’t processing any messages), the health perform creates a set of buyer keys from the buyer service and a set of buyer overseas keys from the ticket service and asserts that the entire ticket overseas keys are contained inside the set of buyer keys.
Why not simply question the databases instantly from the health perform? Abstracting them as units permits flexibility—querying throughout databases on a relentless foundation introduces overhead that will have unfavorable uncomfortable side effects. Abstracting the health perform examine from the mechanics of how the information is saved to an summary knowledge construction has at the least a few benefits. First, utilizing units permits architects to cache nonvolatile knowledge (like buyer keys), avoiding fixed querying of the database. Many options exist for write-through caches within the uncommon occasion we do add a buyer. Second, utilizing units of keys abstracts us from precise knowledge gadgets. Information engineers want artificial keys to utilizing area knowledge; the identical is true for architects. Whereas the database schema may change over time, the workforce will all the time want the connection between clients and tickets, which this health perform validates in an summary approach.
Who executes this code? As this downside is typical in distributed architectures resembling microservices, the widespread place to execute this governance code is inside the service mesh of the microservices structure. Service mesh is a common sample for dealing with operational considerations in microservices, resembling logging, monitoring, naming, service discovery, and different nondomain considerations. In mature microservices ecosystems, the service mesh additionally acts as a governance mesh, making use of health features and different guidelines at runtime.
This can be a widespread approach that architects on the utility stage can validate knowledge integrity, and we’ve applied most of these health features on a whole lot of tasks. Nevertheless, the specificity of the implementation particulars makes it tough to broaden the scope of most of these health features to the enterprise architect stage as a result of they embrace too many implementation particulars about how the undertaking works.
Brittleness for metadomains
One of many key classes from domain-driven design was the thought of retaining implementation particulars as tightly certain as doable, utilizing anticorruption layers to stop integration factors from understanding too many particulars. Architects have embraced this philosophy in architectures like microservices.
But we see the identical downside right here on the metalevel, the place enterprise architects wish to broadly management considerations like knowledge integrity but are hampered by the space and specificity of the governance requirement. Distance refers back to the scope of the exercise. Whereas utility and integration architects have a slim scope of duty, enterprise architects by their nature sit on the enterprise stage. Thus, for an enterprise architect to implement governance resembling referential integrity requires them to know too many particular particulars about how the workforce has applied the undertaking.
Considered one of our largest international purchasers has a task inside their enterprise structure group referred to as evolutionary architect, whose job is to determine international governance considerations, and now we have different purchasers who’ve tried to implement this stage of holistic governance with their enterprise architects. Nevertheless, the brittleness defeats these efforts: As quickly because the workforce wants to alter an implementation element, the health perform breaks. Although we regularly sofa health features as “unit exams for structure,” in actuality, they break a lot much less typically than unit exams. (How typically do adjustments have an effect on some basic architectural concern versus a change to the area?) Nevertheless, by exposing implementation particulars outdoors the undertaking to enterprise architects, these health features do break sufficient to restrict their worth.
We’ve tried a wide range of anticorruption layers for metaconcerns, however generative AI and MCP have supplied the perfect answer thus far.
MCP and Agentic Governance
MCP defines a common integration layer for brokers to question and devour capabilities inside a specific metascope. For instance, groups can arrange an MCP server on the utility or integration structure stage to reveal instruments and knowledge sources to AI brokers. This gives the proper anticorruption layer for enterprise architects to state the intent of governance with out counting on implementation particulars.
This enables groups to implement the kind of governance that the strategically minded enterprise architects need however create a stage of indirection for the small print. For instance, see the up to date referential integrity examine illustrated in Determine 3.
In Determine 3, the enterprise architect points the overall request to validate referential integrity to the MCP server for the undertaking. It in flip exposes health features through instruments (or knowledge sources resembling log information) to hold out the request.
By creating an anticorruption layer between the undertaking particulars and enterprise architect, we are able to use MCP to deal with implementation particulars in order that when the undertaking evolves sooner or later, it doesn’t break the governance due to brittleness, as proven in Determine 4.
In Determine 4, the enterprise architect concern (validate referential integrity) hasn’t modified, however the undertaking particulars have. The workforce added one other service for consultants, who work on tickets, that means we now must validate integrity throughout three databases. The workforce adjustments the interior MCP device that implements the health perform, and the enterprise architect request stays the identical.
This enables enterprise architects to successfully state governance intent with out diving into implementation particulars, eradicating the brittleness of far-reaching health features and enabling rather more proactive holistic governance by architects in any respect ranges.
Defining the Intersections of Structure
In Structure as Code, we focus on 9 totally different intersections with software program structure and different components of the software program improvement ecosystem (knowledge representing one among them), all expressed as architectural health features (the “code” a part of structure as code). In defining the intersection of structure and enterprise architect, we are able to use MCP and brokers to state intent holistically, deferring the precise particulars to particular person tasks and ecosystems. This solves one of many nagging issues for enterprise architects who need to construct extra automated suggestions loops inside their techniques.
MCP is nearly ideally fitted to this objective, designed to reveal instruments, knowledge sources, and immediate libraries to exterior contexts outdoors a specific undertaking area. This enables enterprise architects to holistically outline broad intent and depart it to groups to implement (and evolve) their options.
X as code (the place X might be all kinds of issues) sometimes arises when the software program improvement ecosystem reaches a sure stage of maturity and automation. Groups tried for years to make infrastructure as code work, but it surely didn’t till instruments resembling Puppet and Chef got here alongside that would allow that functionality. The identical is true with different “as code” initiatives (safety, coverage, and so forth): The ecosystem wants to supply instruments and frameworks to permit it to work. Now, with the mixture of highly effective health perform libraries for all kinds of platforms and ecosystem improvements resembling MCP and agentic AI, structure itself has sufficient assist to hitch the “as code” communities.
Be taught extra about how AI is reshaping enterprise structure on the Software program Structure Superstream on December 9. Be a part of host Neal Ford and a lineup of consultants together with Metro Financial institution’s Anjali Jain and Philip O’Shaughnessy, Vercel’s Dom Sipowicz, Intel’s Brian Rogers, Microsoft’s Ron Abellera, and Equal Consultants’ Lewis Crawford to listen to hard-won insights about constructing adaptive, AI-ready architectures that assist steady innovation, guarantee governance and safety, and align seamlessly with enterprise targets.
O’Reilly members can register right here. Not a member? Join a 10-day free trial earlier than the occasion to attend—and discover all the opposite assets on O’Reilly.