🧭 You removed the gate. What replaced it?


Sep 27, 2026 | Read time: 14 mins!

Kevin Donovan · Architects' Hub

Hey Reader and welcome to the Hub.

There’s a moment in the move from gates to guardrails that deserves more attention.

The review meeting disappears. Teams no longer wait for a signature. The calendar looks better.

Then someone asks who is catching the shared dependency, the conflicting data definitions, or the decision 3 teams are making separately.

The meeting used to make those things visible. Removing it doesn’t decide what happens to them next.

This week, I’ve been working through that gap. What does the architect actually do, and what must a guardrail contain before a team can use it?

In this edition

  • The work that survives the review board.
  • The 5 fields, and the failure each prevents.
  • 3 worked examples: auth, a contested data boundary, and an unowned library.
  • A Monday rewrite and a quarterly check.
  • The free 5-Field Guardrail sheet.
  • Community Corner, with the questions from Nadeem and Paul that started this week.

Let’s get into it.👇

________

🧭

________

​

The decision still needs a home

Imagine a review board that catches 2 teams building different versions of the same integration.

The board is slow. The teams queue. Someone quite reasonably removes the approval step.

Both teams can now move faster. They can also finish both integrations before anyone notices the duplication.

The decision still exists. It needs a place where the teams can see it early enough to act.

That might be a shared pattern in a starter repo, a named owner for the integration, or a conversation before either team commits.

The architect’s work shifts toward those earlier moments. Coach the trade-off. Name the dependency nobody owns. Connect teams making the same decision. Keep the guardrail understood.

The evidence changes too. Counting reviews tells you how often people came through the old process. It says little about whether the new arrangement helps.

Start with one real decision. Record when you became involved, what changed, and what the team could do afterward.

“Connected 2 teams before they built separate interfaces” is useful evidence. “Attended 6 alignment meetings” leaves the result unclear.

________

🧭

________

​

Five fields that make a guardrail usable

Here’s the Tuesday Afternoon Test:

Can a team apply this on a Tuesday afternoon, under deadline, without booking time with you?

“Prefer managed services” tells a team what you value. It leaves them to discover which service, for which workload, with which exceptions.

A usable guardrail carries enough of the decision for the team to act.

TABLE HERE??

The boundary deserves the most care.

“Use this for internal services” sounds clear until someone asks about a supplier-hosted service, a reporting pipeline, or an acquired business.

Those questions tell you where the decision is incomplete. Write down the exclusions as carefully as the included cases.

If the boundary is contested, say so. Name the people who can settle the contested part, the temporary behavior, and the review date.

That is more useful than a confident sentence that hides an unresolved disagreement.

The escape hatch also needs a real agreement behind it. Tell teams which deviations they can make and record without waiting. Keep mandatory controls explicit.

Changing the heading to “convention” does not, by itself, grant anyone decision rights.

On LinkedIn, Karol Lacuska put the underlying question better than I did this week. It comes down to mandate: what the board decides, what the teams decide, and what stops people quietly routing around a pattern.
​
My answer in the comments was that teams own their delivery decisions, and the board owns the shared boundaries: data, security, and anything the next team has to live with.
​
Going around a pattern should be allowed, but visible. An exception costs a short explanation. A silent workaround should cost more than that.

________

🧭

________

​

Three worked guardrails

These are illustrative patterns you can adapt. The owners and timeboxes below are example assignments, not claims about a particular client.

1. Service-to-service authentication

Start with a case where the boundary is reasonably clear.

Decision Internal service-to-service calls inside the platform VPC use the platform mTLS pattern.

Boundary This covers services operated on that platform. Partner-facing APIs and data-platform workloads use their separately defined patterns.

Default The platform team maintains a starter repo that includes the configuration. A team following the normal setup gets the agreed pattern without requesting it.

Escape hatch Within the agreed exception scope, the delivery team records the reason, alternative, owner, and review date in its ADR. It tells the platform channel. Cases outside that scope go to the named platform owner for a decision.

Expiry trigger Review the pattern when the service-mesh migration starts or the identity model changes. The platform owner owns that review.

The test is practical. Ask a team to start a service using the template. If it must contact the architect to finish the setup, the default needs work.

2. A data boundary nobody fully agrees on

Customer data is harder. Sales owns the commercial relationship. Finance owns billing details. Both may reasonably describe their record as “the customer.”

A guardrail declaring one system the owner of everything would settle the sentence while leaving the disagreement intact.

Decision Customer relationship attributes come from CRM. Billing identity and invoice terms come from finance. Consuming teams use the published interface for each agreed field.

Boundary This covers the fields in a shared ownership map. Account hierarchy and legal-entity matching remain contested and are explicitly excluded from the settled boundary.

Default Existing integrations keep their current mapping for the contested fields. New consumers use an isolated adapter with a documented mapping. No team silently promotes its local interpretation into the shared contract.

Escape hatch A delivery team with a deadline records the disputed field, temporary mapping, affected consumers, and rollback method. The sales data owner, finance data owner, and consuming team meet to decide the contested case. In this example, they commit to responding within 2 working days.

Expiry trigger The temporary mapping expires after 30 days, or sooner if ownership is agreed. One named data owner tracks that date and brings unresolved cases back to the people with decision rights.

This contains a decision meeting. It is confined to a known unresolved boundary, with named participants and a response time.

If every new consumer needs that meeting, the settled boundary is too small or too vague. Revise it. Do not turn a temporary arrangement into a permanent review board.

3. The shared library nobody owns

An unowned library needs an honest transition plan. Asking every team to “use the shared component” adds dependency without support.

Decision Freeze new adoption of the legacy library. Move new work to the supported replacement.

Boundary Existing consumers can continue on the pinned version during migration. New services are outside that allowance. Necessary fixes have a separate, named maintenance route.

Default Starter repos use the replacement. The delivery pipeline flags new legacy dependencies. An inventory names each existing consumer and its migration owner.

Escape hatch A team unable to migrate records the blocker, temporary containment, owner, and review date. It notifies the interim maintainer and affected consumers. The interim maintainer owns triage of necessary fixes.

Expiry trigger Review when the last consumer migrates or after 60 days, whichever comes first. If the replacement cannot support a required use case, revise the migration plan.

Before adopting this guardrail, leadership must assign an interim maintainer and capacity. A document cannot supply either. Without them, the transition is still an unresolved ownership decision.

________

🧭

________

​

Try this Monday: one guardrail, 20 minutes

Pick a standard that teams repeatedly ask you to interpret. Bring one person who uses it in delivery.

  1. First 5 minutes: write the decision and its boundary. Include at least one exclusion.
  2. Next 5 minutes: find the actual default. Open the template, pipeline, or starter repo. If nothing implements it, name who will build that path.
  3. Next 5 minutes: write the escape hatch and expiry trigger. Give the review to a named person.
  4. Last 5 minutes: ask the delivery colleague to apply it to a recent case without your explanation. Mark every question they still need answered.

Leave with one owner, one change to make, and a date to try it with a team. An incomplete field is a useful finding.

For the next 2 weeks, keep a small decision log. Record whether a team used the default, reported a deviation, or needed help interpreting the boundary. Note any waiting time.

Use the result to change the guardrail:

  • Repeated interpretation questions mean the wording or boundary needs revision.
  • Repeated deviations for the same reason suggest the default needs revision.
  • Waiting for the architect on an in-scope case means the approval queue has returned.
  • A change that contradicts the decision’s assumptions triggers a review. Retire the guardrail if the decision is no longer needed.

These observations are more useful than a compliance percentage with no explanation of the exceptions.
​
​Simon Wright pushed this one step further this week on LinkedIn, and it changed how I think about the log. Exceptions are a tell. Sometimes the team needs help. Sometimes the guardrail needs an update.

If 3 teams hit the same exception, you probably have your next guardrail.

That also answers what the review board is for once it stops checking every design. Its job becomes the cases that stepped outside, because that is where the next version of the rule is already forming.

Once a quarter, explain it again

Book 20 minutes with a team for each guardrail set. Use a recent decision and ask:

  • What does this stop you doing?
  • When did you last work around it?
  • What would make it wrong now?

Ask for the example behind each answer. No reported workarounds could mean the default works well. It could also mean people do not expect a useful response.

The guardrail owner records one outcome: keep it, revise it, or retire it. Assign a date to any revision and bring the result back to the team.

________

🧭

________

​

Your free 5-Field Guardrail sheet

The 4-page sheet brings the method together:

  • The Tuesday Afternoon Test.
  • The 5 fields and the failure each prevents.
  • One worked auth guardrail and 2 blank spaces for your own.
  • The 3 upkeep questions and guidance for a contested boundary.

​Download The 5-Field Guardrail​

Free, no form or checkout. Pick one guardrail and use the sheet with the team that has to apply it.

________

🧭

________

​

Community Corner

🧭 Community Corner

This one comes from LinkedIn rather than the inbox, which is a first.

Usually this section is built from your replies to this email. This week a bunch of you got into the comments instead, and several of you made the argument better than my post did. So I am putting four of those in front of everyone, with names on them.

1. A second test, and a sharper one than mine

​Pedro Anastácio, a solution architect, read the Tuesday Afternoon Test and immediately found the softer version of the problem I was describing:

“A lot of ‘guardrails’ are really just approved patterns written down somewhere, which is a softer gate, still depending on someone finding it, reading it, and remembering it at the right moment.”

That is uncomfortable, because it describes most of what gets called a guardrail, including some of mine. A pattern in a wiki is not a guardrail. It is a gate that has stopped staffing itself.

His alternative is to stop relying on people remembering:

“The guardrails that actually hold are built into the system itself: a typed interface that makes the wrong integration path harder than the right one, or a shared infra layer that centralizes retries and error contracts. Those don’t need to be discovered, they’re just the path of least resistance.”

Then he hands you a test I wish I had written:

“For each guardrail, ask whether it’s a constraint or a recommendation. If it can be skipped without anyone noticing, it isn’t a guardrail yet.”

Put that next to field 3. I wrote that the default should be the compliant path. Pedro gives that a pass and a fail. If someone can quietly take the other road and nothing anywhere registers it, you have written a recommendation and filed it under governance.

My one qualification is about sequence, not substance. Nothing gets built into the platform on day one. Most guardrails start as a written pattern, and the good ones earn their way into a typed interface or a shared service once enough teams lean on them. Written down is the first step. Built in is where you are aiming.

2. Do AI agents make the review board unnecessary?

​Trevor Allen, who runs an architecture practice, made the case for moving accountability back where it belongs:

“It’s not enough to build the guardrails, your architects need to be fully aware of them and why they exist… move the accountability onto the architect to adhere to them, and not hand the accountability over to the ARB because they voted that the design was ok.”

Agreed, and firmly. A board vote is not a place for ownership to go and hide. Plenty of designs have been waved through by people with less context than the person who wrote them, and the vote changed nothing except who felt responsible afterward.

Then he goes somewhere I cannot quite follow:

“No single architect can ever keep all guardrails in mind, so the value of an ARB or peer review panel is to give assurance for the harder to grasp and nuanced guardrails. This is where AI guardrail agents will become valuable and remove the need for the review board.”

The first half is obviously true. Nobody holds 40 guardrails in their head, and pretending otherwise is how the quiet exceptions start. And I think he is right that agents will take the routine checks, which is a good use of them: let the machine catch the boring violations early so people spend their review time on judgment.

Where I part company is the last clause. An agent can tell you a guardrail was crossed. It cannot tell you the guardrail is wrong. Someone still has to look at an exception and decide whether the team needs help or the rule needs retiring, and that decision is the actual work. Take the board’s checklist away and I think it survives. Take away its judgment and you have automated the gate rather than replaced it.

I would rather be wrong about this than right. If your org has an agent doing more than routine checking, I want to hear about it.

3. The line I am borrowing

​Christoffer Råsten, 20 years across insurance and media, wrote one sentence I have not stopped thinking about:

“Architecture is not a place, it’s a competence.”

He follows it with the failure mode this whole edition is about:

“Guardrails evolve over time. As teams, technology and context change, they need to evolve too. Keeping them visible, owned and easy to challenge helps them remain enablers of good decisions rather than gradually becoming gates themselves.”

Easy to challenge is the part I would underline. A guardrail nobody questions has already turned back into a gate, it just has not been noticed yet. That is what field 5 is for, and Christoffer got there on his own without needing the field.

4. Nobody talks about how you actually build them

​Jordan Corn asked the question that the entire week skated past:

“In all the discussions I’ve heard about guardrails, almost nobody talks about the challenges of building them and then driving their adoption. If the guardrails aren’t enforced, and don’t provide a faster path to implementation than going it alone, it seems to me they’re unlikely to succeed.”

He is right, and this edition is guilty too. I gave you 5 fields and a test and said almost nothing about getting anyone to use the thing.

The short version of what I have seen work is speed. When the approved auth pattern ships as a template that puts a team in production in a day, nobody argues with it, because arguing costs more than complying. When the guardrail is slower than going it alone, teams go it alone. Every time, and they are right to. Jordan even made his own infographic for his teams:

But Jordan did not want my answer. He wanted yours:

“I’d love to hear some… any… stories about how people made them work AND made implementation faster!”

So that is this week’s ask, and it is his, not mine. If you have made a guardrail that teams actually reach for because it is the fast path, hit reply and tell me how. I will put the best of them in front of everyone in a fortnight, with your name on it if you want it there.

Thank you all for being part of Architects’ Hub.

Kevin

________

________

Kevin Donovan​
I write Architects' Hub every Sunday for senior architects and technology leaders.
You're getting this because you subscribed at architects-hub.kit.com.
Know someone who’d find this useful? Forward it to them.

​

Korsvejen 9, Viborg, Viborg 8800
​Unsubscribe · Preferences​

Architects' Hub

A Sunday newsletter for senior and enterprise architects. One decision, one practical model, one tool you can use Monday. Written by Kevin Donovan. Join 2,500+ architects. One email every Sunday.

Read more from Architects' Hub
Three doors slightly ajar in a row. Through them, a telephone, a stack of documents with a magnifying glass, and an arrow curving back on itself. A single coin rests on the floor in front of the doors.

Nov. 22 2026 | Read time: 11 mins. Kevin Donovan · Architects' HubForwarded this email? Subscribe to Architects’ Hub for the next issue. The room went quiet Hey , Two Sundays ago, our hypothetical pilot case got its commitment record. Last Sunday it got a walkthrough, and the operating owner signed with 3 tests on record. The AI assistant now drafts delayed-order replies for 3 teams, and the drafts are good. This week someone proposed letting it do more than just draft… In this edition The...

Five simple figures in a ring around a small central box, each holding one object: a magnifying glass, a signpost, a wrench, a mop, a headset. One figure carries a leaning stack of boxes.

Nov 15, 2026 | Read time: 11 mins. Kevin Donovan · Architects' HubForwarded this email? Subscribe to Architects’ Hub for the next issue. Tuesday, 10:40 am. Hey ,Last Sunday we left a hypothetical customer-service assistant with a commitment record and unanswered questions about capacity, data and support. This week, move the story forward 3 weeks. The team has agreed to a bounded trial on 1 queue. Agents still review every draft. Wider rollout is waiting. I want to spend this edition with the...

A working terminal with four loose connections labeled Funding, Operations, Data and Authority.

November 8, 2026 | Read time: 9 mins. Kevin Donovan · Architects' HubForwarded this email? Subscribe to Architects’ Hub for the next issue. Hey ,In this edition of Architects' Hub: What the service lead heard when the sponsor said, “We should roll this out.” A commitment record filled in for one example customer-service queue, PLUS a blank copy for yours. How to take a conditional decision back to the sponsor who wants a date. Community Corner: Santiago Jesus Lasso Valenciano on the work that...