A manifesto for founders who want more than a faster code factory
You can run a startup today where everyone uses AI and the company is still organized as if it were 2019.
The developers work faster. Marketing produces more. Sales has automatically written emails. The CEO gets summaries.
And the customer is still waiting.
Every department has been given a turbocharger. The gearbox is still lying in the car park.
That is where AI Ops begins for me: continuously designing the information, decision and verification flows through which people, agents and software deliver reliable outcomes together.
Not as another IT project. As a core part of strategy.
Here, AI Ops means organizing AI-enabled work across the entire business, rather than the narrower use of AI for IT operations. FluxWorks is fictional. Its examples illustrate design choices, not measured results.
At the heart of AI Ops is continuous deployment applied to the organization itself: how you work and the strategy you pursue. Not by setting an AI strategy once, but by continually testing and revising how we work and what we choose to pursue. People, persistent assistants and software factories are useful combinations today. They are examples of what is possible now, not the end state. The shape of AI Ops will change as AI technology changes. What lasts is the ability to keep testing, deploying and rolling back improvements to how we work and to our strategy.

TWO LEARNING LOOPS
Improve the work. Redesign how work happens.
Outer loop · leadership
- Choose goals and limits
- Design roles, mandates and process
- Review daily evidence
- Decide redesign ↺
Inner loop · daily work
- Observe
- Decide
- Act
- Verify
- Learn
Daily evidence informs redesign. New goals and limits guide the next daily cycle.
The daily loop delivers and learns. In the leadership loop, that evidence informs small, reversible releases of how you work and which strategy you pursue.
FLUXWORKS · FICTIONAL
FluxWorks: from customer request to a feasible schedule proposal
Fully fictional design example, not a customer case or working implementation. All counts, timings, budgets and outcomes below are illustrative assumptions, not measurements.
- Customer outcome and owner
- FluxWorks helps a manufacturer determine a feasible delivery date for 120 housings. Maya, operations lead, owns the flow from request to accepted delivery promise, not just machine utilization.
- Baseline — illustrative
- In a fictional reference set of 20 requests: median 2 working days to an approved proposal, 4 proposals required revision and 16 of 20 delivery promises were met. These are chosen starting values for this example, not collected data.
- Suspected constraint
- Sales, planning and procurement use different versions of capacity and material availability. Most time is lost at handoffs. Hypothesis: one shared, traceable planning context helps more than faster planning inside a single department.
- Daily loop
- Maya’s persistent assistant reconciles a changed customer request in email with the calendar, recent project decisions, and approved capacity, stock and maintenance snapshots. It retains the version and timestamp for each source. The planner chooses two scenarios. If the checks need to change, the assistant helps Maya brief the software factory with a bounded task and acceptance tests. The factory builds and tests outside production. The assistant only drafts the proposal. Hard checks cover materials, capacity, sequence and existing customer commitments. Maya releases the delivery promise. Exceptions become test cases.
- Agreed acceptance criteria
- Replay the same 20 fictional requests without changing production. No capacity conflicts, material shortages or worse existing delivery promises. Every proposed date must trace to its inputs; missing data block release. Illustrative target: median at most 1 working day and at most 2 revisions, with no worse delivery reliability. No success claim before evaluation.
- Mandate and budget
- Read only from the approved test set, at most two scenarios per request, at most €25 compute spend and 2 hours of human review for this experiment. No ERP writes, customer messages or machine control. More budget or authority requires Maya’s decision.
- Evidence and independent challenge
- Retain input versions, scenarios, check results, reviewer objections and Maya’s acceptance for each request. Compare with the old approach on the same dataset. A second agent alone is not independent evidence: hard checks and human-selected exception cases remain necessary.
- Stop rule and fallback
- Stop for missing or contradictory inputs, a violated hard constraint, exhausted budget or an unexplained deviation. No automatic release. The existing planning process remains authoritative; Maya handles the exception.
- Open track: let a few ants wander
- Alongside the planned trial, a small group from sales, planning and procurement gets protected time to choose what to explore. No business case up front, but clear time and data limits and no production permissions. With a pirate flag in the corner, they try something different: could the sales-to-planning handoff disappear if both teams used the same traceable request record? Using fictional requests, they show what works and where a manual handoff is still needed.
- Transfer: does it work without the pioneers?
- A colleague outside the group tries the new approach on their own practice requests without the creators taking over. Compare time to an accepted proposal, including review, rework and help, against the existing process. Check whether the colleague chooses to use it again on the next request. If it only works in the demo, simplify the workflow or drop it.
- Leadership loop
- Maya and the founder review weekly where waiting and errors actually arise. If unclear ownership is the constraint, change the handoff or decision rights, not just the prompt. Test new process versions alongside the current process first.
- Decision after the experiment
- If the evidence improves, Maya gives one team a limited, reversible mandate. The old handoff stays available if capacity, exception handling or delivery reliability deteriorates. Maya and the founder then decide both which workflow to keep and where to put people and budget. They might stop speeding up the old handoff and invest in the shared request record instead. A persistent assistant and a software factory help today. The tools may change; the loop of discovery, testing, rollout and rollback remains.
Agent proposes. Tests challenge. Owner accepts.
1. Start with the customer, not your agent diagram
Imagine FluxWorks, a fictional AI startup that helps manufacturers build better production schedules. Eight people, a growing flock of agents and a founder who thought less programming would automatically mean less work.
That was a lovely thought.
Sales promises a new feature. Engineering builds it. An agent optimizes machine utilization. Customer success discovers that the customer mainly wanted reliable delivery dates.
Everyone has done their job. The company has missed the problem.
So the first question is not: how many agents do we need?
It is: what outcome does the customer want, what must we know and decide to deliver it, and what evidence will show that it works?
Follow one customer request through the whole organization. Especially through the awkward gaps between departments. That is often where the queues, lost context and unanswered questions live.
2. Organize around outcomes and give someone ownership
At FluxWorks, ‘a reliable production schedule for this customer’ has one human owner. Not because that person does everything, but because someone must understand the whole and remain accountable.
Agents analyze, build, simulate, check and document. The owner defines the goal, the mandate and when the evidence is insufficient.
Give every important value stream a working agreement: the desired outcome, relevant context and sources, permitted actions, quality and stop criteria, budget, owner and escalation route.
An agent with a job title is not an organizational design.
3. Communication is infrastructure
A Markdown file is text. A skill transfers knowledge and ways of working. Project context explains what matters. The value is not in the filename. It is in the information, its relevance and how it is used.
At FluxWorks, a customer promise is connected to the product decision, acceptance test and delivery status. Facts, assumptions and outdated information are distinguishable.
More context does not automatically mean better context. A moving box full of meeting notes is information too. Good luck finding Tuesday’s agreement.
Good communication gets the right information to the right decision-maker at the right time. With its source, uncertainty and context.
A perfectly worded prompt will not repair an organization with three different definitions of success.
4. Automate the evidence too
FluxWorks asks an agent to improve its scheduling algorithm. Average machine utilization rises. Excellent.
Then an important customer order is delivered late.
The test confirmed exactly what had been requested. Unfortunately, the wrong thing had been requested.
Define acceptance criteria before starting the optimization loop: delivery reliability, cost, safety, capacity and exceptions. Use hard checks wherever possible. Use agents for interpretation, discovering missing scenarios and investigating deviations.
StrongDM’s software factory is interesting because human attention shifts from every line of code towards specifications, scenarios and test environments. That is one company’s approach, not proof that the same recipe is safe everywhere.
And test the testing. Two agents sharing the same mistaken assumption make a very orderly meeting. Not necessarily independent evidence.

5. Autonomy must be earned
How much can AI do independently? As much as the process can demonstrably support. Not as much as makes the demo impressive.
An agent at FluxWorks may simulate thousands of schedules independently. It may not casually change a delivery commitment or operate a real machine.
Build autonomy progressively: evidence-backed proposals → testing in an isolated environment → reversible actions within a mandate → continuous execution with monitoring, budget limits and escalation. Expand that mandate only when both quality and recovery have been demonstrated.
Human review belongs where uncertainty and consequences meet. Not as a ceremonial click after every action.
A human approving a hundred proposals an hour is not a safety mechanism. That is a tired human.
6. Efficiency, redundancy and adaptability
Efficiency helps FluxWorks deliver value faster. Redundancy prevents one failed provider, mistaken assumption or lost piece of context from bringing the company to a halt. Adaptability allows the company to change how it works as capabilities change.
These three sometimes pull in different directions. A second check takes time. An alternative model route requires maintenance. Experimentation looks like unused capacity in the short term. Until you need it.
Redundancy does not mean duplicating everything. It means deliberately choosing where independent verification, spare capacity and recovery paths justify their cost.
And adaptability does not mean replacing your entire stack every Thursday because somebody on LinkedIn typed ‘game changer.’
Change because a better way of working is demonstrably available.
7. A flock is not a strategy
Sometimes one model with good tools is enough. Sometimes a fixed workflow is better. Sometimes parallel research by multiple agents genuinely adds value.
Google describes its /swarm preview, announced on 10 November 2026, as a way for a coordinator, specialist agents and verifiers to divide work, communicate and combine results. Interesting. But a preview is not evidence that more agents are better for every task.
Additional agents help only when the work can be meaningfully divided and communication, integration and verification do not consume the gains.
At FluxWorks, agents investigate different approaches. A coordinator maintains the shared question. An independent evaluation compares the results against predefined criteria.
They do not all need to read everything from everyone else. Organizations have been drowning in cc’d emails for long enough. We do not need to pass that tradition on to our software.
An assistant that still knows why tomorrow
AI is not just a specialist for a single workflow. One practical way to use it broadly is a persistent assistant: a generalist that can carry agreed goals, preferences, decisions and open work across conversations and tasks. At FluxWorks, Maya and her assistant compare a customer email with the commitments in her calendar and the latest project decisions. Together they brief the software factory and define acceptance tests. The factory builds and tests, the assistant keeps the work aligned, and Maya remains accountable. The assistant can support one person or be shared across a team.
Broad usefulness does not mean broad authority. For each task, the assistant retrieves only the relevant, permitted pieces of personal, team or organizational context: a customer commitment, a decision or a working agreement, with its source, date and owner. Access stays limited and revocable. Preparing something is not the same as sending, approving or deploying it.
8. Optimize the whole flow, not the busiest component
If building goes from ten days to one, but a customer decision takes three weeks, the next bottleneck probably is not a programming problem.
Measure the time from customer request to accepted outcome. Include waiting, rework and decision-making. Measure quality, cost per accepted outcome, recovery time and human attention.
Tokens, commits and agents tell you something about activity. Not necessarily value.
Every acceleration raises a new question: what now limits the whole? A successful improvement moves the bottleneck. It does not abolish it forever.
Lean, DevOps and MLOps have therefore not become obsolete. They become relevant again: flow, feedback and reliability, now with new participants in the system.
9. Organize your own creative destruction
FluxWorks’ founders have two responsibilities: keep the company working and investigate which parts of that same company should work differently.
Which handover could disappear? Which task exists only because two systems do not understand each other? Which interface is needed for human review, and what can an agent do directly through a tool?
Treat processes as temporary designs. Experiment with new models, different divisions of work and sometimes an entirely new process. Compare them against the same customer outcome, risks and costs.
But replace them in a controlled way: test first, then run alongside the existing process without taking over decisions, then deploy within a limited scope. Expand only with evidence and a fallback.
Creative destruction without a recovery plan is usually called an outage in a production environment.
Let a few ants stray from the trail
Pick one bottleneck to start on Monday. Also protect time for a few people to explore which work could be done differently or removed altogether, without making them prove a business case up front. That is the wandering ant principle: most ants follow the known trail, while a few look where there is no trail yet. Put a pirate flag in the corner if it helps. Challenge the way of working, not the people. Experiment outside production, within clear limits on time, data and authority. ‘This didn’t work here, and here’s why’ is a useful outcome too.

10. Take people seriously
People are often given a new tool and then left with the same targets, the same meetings and the same performance reviews. We call their frustration resistance to change.
Let people surface the exceptions, failures and unwritten agreements. Give them room to practise. Make it clear who decides, how quality is assessed and what changing roles will mean.
Not everyone needs the same technical role. But people must be able to report relevant failures, question the system and intervene within their responsibilities.
Some roles will change profoundly or disappear. Be honest about that. But do not treat people like an outdated software library either. Fair decisions and proper support are part of good management.
11. Learn from evolution without copying its cruelty
Variation → selection → retention. Try different ways of working. Test them against reality. Keep what works. Remove what no longer helps.
Without reliable selection, you scale mistakes just as quickly. And an organization must decide for itself which outcomes are desirable and which harms are unacceptable.
I call the practical design question fittest affordance: which combination of person, model, tool and context best fits this task right now, within our time, resources and constraints?
The most powerful solution does not automatically win. Neither does the cheapest. The best-fitting combination wins. Until the circumstances change.
12. Make learning cumulative
A failure becomes a test. A customer conversation changes a hypothesis. An incident clarifies a boundary. A successful exception becomes standard practice only when we understand why it worked.
Preserve provenance, versions and the conditions under which knowledge remains valid. And forget deliberately: an outdated instruction is not intellectual capital just because it lives in a folder.
A demo shows what is possible. A colleague shows whether it transfers.
Innovation adoption matters even when you are your own customer. Let pioneers demonstrate real work, including the checks and the mistakes. That helps colleagues see what is possible. Then ask someone outside the pilot to repeat the workflow with a case of their own. Does it work without its creator? Is the result accepted sooner once you count review and rework? Only then give the workflow an owner, instructions and a fallback. Help pioneers pass on what they learned instead of turning them into the permanent help desk.
THE MANIFESTO
We organize around customer outcomes.
We make intent, context and ownership explicit.
We automate verification alongside production.
We expand autonomy where evidence and consequences allow it.
We combine efficiency with appropriate redundancy and adaptability.
We measure the entire value stream.
We treat processes as designs that can be replaced.
We give people agency, responsibility and room to learn.
We make room for targeted improvement and open exploration.
We put proven improvements into practice in how we work and in our strategy.
AI Ops is not an implementation you eventually finish. It is the ability to keep today’s company working while building tomorrow’s.
Apply continuous deployment to the organization itself. Make small, reversible releases of your own way of working: test them alongside the existing approach, measure the outcome, roll them out deliberately and roll them back if results get worse. Sometimes you change a handoff or a tool. Sometimes you change strategy: which initiatives you stop or fund further, where you put people’s time, or which authority you extend or withdraw. Let the evidence inform the next decision. ‘Continuous’ means the learning never stops, not that an agent automatically changes production.
Accelerate with control.
And occasionally have the courage to admit that you need a different car.
From manifesto to first experiment
A possible first month for FluxWorks, offered as a working hypothesis rather than a universal recipe:
Work on two tracks. Measure one existing customer value stream and test a targeted improvement alongside the old approach. At the same time, protect a recurring block for open exploration: let a small group explore other work, share what they learn and ask a colleague to repeat a promising trial. Review both tracks weekly. Decide explicitly what to stop, revise, roll out on a limited basis or roll back, and whether the evidence warrants changing priorities, resources or authority.
- Choose one customer value stream and establish a baseline for lead time, quality, cost, rework and human review time. Assign an owner.
- Make the goal, context, mandate, acceptance tests, budget and stop mechanism explicit. Define who has decision-making authority.
- Run the new approach alongside the existing one without autonomous production actions. Compare outcomes and exceptions.
- Grant a limited, reversible mandate when the evidence is strong enough. Review weekly, expand, repair or roll back.
Keep two loops visible: the daily loop of observe → decide → act → verify → learn, and the leadership loop that redesigns the way of working itself.
STARTER KIT · ENGLISH · V1.0
Make it practical: the Founder’s AI Ops Starter Harness
A blank, editable Markdown starter with working agreements, mandates, evidence records and dry-run tests. Version 1.0 · 15 November 2026. Behavioral instructions and templates, not executable software or a guaranteed security boundary.
- Make a private working copy; never put company data in a public copy.
- Fill in owner, sources, mandate, budget and stop conditions. Unknown stays UNSET.
- Start with discovery and proposals. Test before granting bounded execution.
This does not automatically give your bot tools, memory, background execution or additional authority.
Evidence and limitations
- StrongDM: Software Factories describes its own approach using agents and external scenarios. Valuable practical experience, not a universally proven recipe.
- Google Antigravity: /swarm describes coordination, parallel work and verifiers. It is a preview; more agents also require coordination and potentially much more compute.
- Anthropic: Multi-agent research found gains on a specific internal evaluation with substantially higher token consumption. Those findings cannot simply be generalized to every company or task.
- OpenAI: Harness engineering illustrates why context, tools, feedback and observability together amount to more than a long prompt.
- DORA 2025 describes AI as an amplifier of existing organizational strengths and weaknesses. It does not establish a universal causal effect for every organization.
- Lean thinking, Conway on communication structures and Google SRE on controlled releases provide older but still relevant foundations.
The definition of AI Ops and its application to FluxWorks above are my synthesis of these sources and our own principles. They are not an established, proven blueprint.
About this version
v1.2 · Updated 20 November 2026. First published and sources checked on 15 November 2026. Original English text and Dutch translation. The AI Ops definition and fittest affordance are my synthesis, not a proven standard. AI-generated illustrations depict fictional FluxWorks, not working software screenshots.
Download the manifesto (Markdown)