I’m penning this in bed following a lengthy day of trying to explain item that seems obvious, but all period I articulate it, it feels vague. I think the current blog part the finest one on explaining what a application factory is! But you inform me…
An AI application factory is a managed scheme (like haze servers are managed hardware) for turning defined requirements into validated application changes using AI agents, connected tools, automated checks, and individual oversight.
If you are comparing the finest AI application factory platforms, do not commencement alongside the figure of agents or the interface. Start alongside whether the scheme connects intent, context, execution, verification, approval, release, and feedback.
Jeez, that was a lot… Let’s clasp our respiration for a small bit more.
It includes the delegate workflow: how activity starts, what environment is available, what actions are permitted, and how results are accepted.
At Camplight, we have spent the final nine months construction application factories and have deployed three for clients.
Those implementations are covered by NDAs. We cannot publish the client systems yet.
Instead, we have generalized our implementation cognition into Nest: a community example that makes the functioning questions and construction blocks apparent without revealing confidential deployments. It’s according to our open-source OrgOps infra.
This is not a prediction concerning item we power eventually build. It is a way to explain activity we are already doing.
The screenshots below current a unified, generalized example fairly than any of the three confidential client environments. They do not connote that all deployment contains all interface shown. I desire to add a disclaimer that the names, budgets, timings, and achievement figures are illustrative, not published client results.
When comparing foremost AI application factory platforms, treat these screens as an evaluation example fairly than a worldwide merchandise checklist. I conjecture I volition have to compose another blog part in brace of months since the area is moving so quick…
The helpful inquiry is not whether your business needs an interface identical to Nest.
It is whether you can answer the operational questions rearward it.
This applicable guide builds on our inspection of Uber’s AI application factory and our inspection of dark-factory readiness. Here, we concentration on what the scheme contains, how activity moves through it, and what leaders need to govern.
What makes an AI application factory distinct from a coding agent?
A coding delegate can execute significant betterment work. For example, GitHub’s Copilot haze agent can examine a repository, scheme changes, modify code, and run tests in a betterment environment. A application factory hence cannot be distinguished merely by saying, “Our AI does additional than autocomplete.”
The difference is the system about the execution.
In the example we use, a application factory connects a endeavor petition to the context, tools, people, validation, and publish procedure needed to provision it.
Defined intent -> applicable environment -> implementation -> verification -> required endorsement -> publish -> feedback
Different parts of that stream can use distinct mechanisms. Some necessitate an agent. Others are improved handled by average code, an existing pipeline, or a person. (hint: normally the flask neck is about verification since evals and guardrails are a moving target)
An AI application factory additionally does not substitute uninterrupted integration and delivery. CI/CD already provides mechanisms for building, testing, and deploying changes; the factory needs to nexus agent-generated activity to those mechanisms.
A coding delegate performs work. A application factory defines how that activity becomes an accepted, accountable application alter that anyone can trigger.
The crucial bit current is “anyone” but we’ll arrive rear to this later.
The term itself is used in distinct ways. Cortex describes an organizational software-delivery system, during StrongDM describes a deliberately non-interactive approach without individual code review. Our range is a human-led AI application factory, alongside definitive decisions concerning autonomy, verification, and intervention. This is since nobody has reached full dreary state
What the finest AI application factories have in common: seven construction blocks
We organized Nest about seven areas that leaders can use to difference AI application factory platforms: operational visibility, delegate management, configuration, projects, squad assembly, reusable capabilities, and individual collaboration.
These are evaluation criteria for a foremost AI application factory, not a necessity to build seven new applications.
1. An operational dashboard: what is happening, and what needs attention?
The dashboard is anywhere a individual have to be capable to comprehend the province of the factory without reconstructing it from conversations. I loved my millennial days alongside mIRC but chatting is excellent tiring.
What is running? What has finished? What is waiting for review? Which issues need intervention? What has implementation cost? asl pls?
Nest brings those questions together through Spend, Tasks, Issues, Active Projects, and Community.
The crucial scheme choice is to nexus visibility to action! A blocked project should guide to its context. A spending anomaly should guide to the applicable workflow. A assessment petition should guide to the relic and the criteria for accepting it.
For an executive, I would additionally differentiate activity metrics from delivery metrics.
“Agents completed 100 tasks” describes activity. It does not established that the business received 100 helpful outcomes.
My preferred evaluation would merge accepted changes, elapsed shipment time, assessment effort, rework, and implementation cost. For comparable work, one helpful measure is:
Cost per accepted alter = total implementation and assessment disbursal for a defined lot of activity / accepted changes in that batch
That calculation should contain unsuccessful attempts and retries, not lone the prosperous final run.
The dashboard additionally includes small celebrations and shared activity from the community. That is deliberate. I desire this surroundings to display what group are accomplishing together, not fair a expanding queue of device activity.
The administration question: Can person see anywhere their notice would be most helpful correct now?
2. Agent management: arrange capabilities about the work
An delegate directory should do additional than catalog names and avatars. It should create responsibilities understandable.
What does all delegate do? Where does it operate? Who owns its configuration? Is it active, paused, or waiting for help?
This is anywhere the inquiry of horizontal versus vertical agents becomes useful, provided we define it.
A horizontal capability power assistance multiple teams, specified as a reusable investigation workflow. A domain-specific capability power run inside one product, repository, or endeavor process.
These are scheme choices, not competing religions.
A shared capability can decrease duplicated work. A capability near to one domain can have a clearer environment and tighter boundaries. The correct choice depends on the activity and the organization.
Team Topologies provides a helpful reference for thinking concerning squad boundaries, shared platforms, and communication modes. Applying those ideas to agents is an architectural analogy, not a assertion that its individual squad types map immediately onto AI roles.
I would not commencement by recreating the company’s org diagram as a gathering of bots.
I would commencement alongside an outcome, acknowledge the capabilities it requires, and decide anywhere those capabilities should live.
The administration question: Does all delegate have a apparent duty and a individual owner?
3. Agent configuration: instructions are lone part of the system
The Agent Details display makes an crucial difference visible: an delegate is additional than its prompt.
In Nest, its configuration includes a clarification of its responsibility, example selection, tool connections, skills, workspace, and additional instructions.
Consider an implementation agent. “Write fine code” is not a adequate functioning instruction.
It needs a defined change, applicable repository context, acceptance criteria, admission to the suitable tools, and an expected output. That output power be a projected code alter accompanied by test results, fairly than approval to publish directly.
There is additionally a border that deserves particular attention:
Instructions depict what an delegate should do. Authorization controls decide what it can do.
“Do not contact production” should not be the lone item preventing manufacturing access. Permissions need implementation exterior the model, alongside admission scoped to the project and high-impact actions topic to suitable approval. OWASP’s agent-security guidance explicitly addresses smallest privilege, tool authorization, and individual endorsement controls.
That is why configuration is an functioning concern, not fair a prompt-writing exercise.
The administration question: What can this delegate admission and change, and who is accountable for that decision?
4. Projects: keep the environment attached to the outcome
The Projects display shows how the pieces arrive together about a particular part of work.
In the illustrated example, a individual asks the squad to validate a pricing strategy. A investigation delegate supplies analysis. A scheme delegate proposes a direction. The individual selects an option. A scheme delegate starts the pricing-page mockup.
The crucial part is not that multiple agents appear in one conversation.
It is that the request, evidence, recommendation, individual decision, and next relic pertain to the same project context.
That creates a clearer handoff than asking all attendee to rebuild the activity from a caller prompt.
The project additionally carries a budget, status, files, and the agents involved. A new attendee have to be capable to comprehend what has been decided and what remains unresolved.
For an genuine application release, the workflow would need to continue beyond the mockup: implementation, testing, review, and the organization’s publish controls. A convincing scheme relic is not the identical item as a validated manufacturing change.
This is what I average by organizing activity about outcomes fairly than secluded conversations.
The administration question: Can you trace a delivered alter rear to the request, evidence, and decisions that produced it?
5. Team assembly: commencement alongside the capabilities the project requires
“Meet your team” is how Nest presents the first gathering of agents.
The screenshot includes strategy, research, product, design, and engineering capabilities. The person can keep or customize the projected team.
This is a way to create configuration approachable. It is not a medication that all workflow needs five agents.
A records update and a new merchandise undertaking should not automatically obtain the identical implementation structure.
Our scheme regulation is to collect the smallest helpful set of capabilities about the problem. Add specialization anywhere it improves quality, control, or clarity, not merely since additional agents appearance impressive.
That is accordant alongside Anthropic’s engineering guidance to commencement alongside the simplest workable resolution and current additional detailed agentic structures lone whenever their benefits validate the additional disbursal and latency.
The avatars assistance differentiate capabilities. They do not transfer accountability distant from the group functioning the system.
The administration question: What does this activity really need, and what complexity can we depart out?
6. Shared skills and automations: create helpful activity reusable
A completed project can create additional than its contiguous deliverable.
It can additionally depart rearward a helpful workflow, connector, education set, or delegate skill.
Nest’s Community area makes those capabilities discoverable. The illustrated examples contain a weekly merchandise beat and a recurring project digest: automations another squad could detect and modify fairly than rebuild.
That is the difference between sharing an answer and sharing the capability to create helpful answers repeatedly.
But reuse needs ownership.
For a shared capability, I would desire to cognize what it does, who maintains it, what inputs it expects, which systems it accesses, and how changes to it are tested. Teams additionally need a way to comprehend whether they are using the identical version.
And sharing a workflow should not average silently sharing the creator’s credentials or admission rights.
Reuse the capability. Reauthorize its access.
The Community display is a apparent expression of this idea: helpful activity have to be capable to journey beyond the individual who archetypal created it.
The administration question: When one squad improves a workflow, can another teams advantage without losing authority of how it runs?
7. Human-agent huddles: create involvement focused and practical
The Huddle display brings group and agents into a focused meeting about the current work.
In the illustrated onboarding redesign, a individual reviews the relic and gives spoken feedback. Speech-to-text turns that feedback into messages the agents can activity with.
The idea is not to put all delegate project inner another meeting.
It is to create an productive involvement item whenever the activity needs clarification, a tradeoff, or a decision.
Imagine a redesign that has passed its checks but motionless misses what the merchandise owner intended. Rather than continuing a sequence of detached corrections, the applicable group can inspect the current artifact, explain the issue, and straightforward the next alter in one session.
The resulting decision should remain attached to the project. A huddle should decrease ambiguity, not create another location anywhere environment disappears.
For consequential actions, spoken feedback additionally needs to remain distinct from authorization. Discussing a deployment is not necessarily approving one.
The administration question: When implementation reaches a judgement call, can the correct group intervene without restarting the work?
What sits underneath the interface: verification and publish controls
A dashboard, delegate roster, and shared conversation do not by themselves create a application factory.
The output motionless needs to be checked.
For application changes, I would anticipate the implementation to define acceptance criteria, run suitable tests and safety checks, keep implementation evidence, and nexus to a controlled publish process. The exact controls should indicate the possible effect of the change.
The delegate reporting “completed” is an event in that process, not the final authority on whether the outcome is acceptable.
StrongDM’s published software-factory work makes this difference particularly concrete: its squad describes using distinct scenarios and simulated assistance environments to measure generated software, including concerns concerning agents satisfying tests without satisfying the intended behavior. That is one implementation approach, not a necessity to copy its complete architecture.
Our human-led framing leaves area for distinct levels of automation.
Some activity can continue through automated checks. Other activity needs an definitive decision from an engineer, merchandise owner, or another accountable person.
Autonomy is a guideline choice for a particular workflow, not a covering commitment that everything should happen without people.
How to commencement construction an AI application factory inner your organization
I would not commencement by commissioning the complete Nest interface.
Start alongside one bounded, recurring software-delivery workflow anywhere achievement can be evaluated.
That power be a well-defined category of care changes, test improvements, or internal-tool updates. Establish how that activity happens today: its volume, elapsed time, assessment effort, and average nonaccomplishment points.
Then define the functioning boundaries before expanding autonomy. What can the delegate read? What may it change? What evidence must it produce? When does it stop? Who accepts the result?
Connect the workflow to the tools you already use anywhere that is practical. A factory does not rotate into additional precious merely since all component is new.
Run a constricted evaluation, inspect the outputs, and difference accepted results alongside the baseline. Expand lone whenever the evidence supports expansion.
This additionally keeps the endeavor conversation concrete.
Which recurring outcome could we provision additional efficiently if execution, context, and verification were connected?
Three deployments in nine months is our implementation experience. It is not a norm shipment timeline or a commitment that all institution volition need the identical scope.
Questions executives ask concerning AI application factories
Is an AI application factory the identical as an AI factory?
Not necessarily. Infrastructure companies specified as NVIDIA use “AI factory” to depict computing infrastructure optimized for AI workloads, including training and inference. This part concerns a distinct subject: the functioning scheme for producing application alongside agents, tools, and people.
Does a application factory substitute engineers?
That is not the example presented here. Engineers remain liable for architecture, specialized standards, verification, and the reliability of the shipment system. We are describing how implementation can be delegated, not how accountability disappears.
Does all business need a multi-agent platform?
No. Our advice is to equivalent the scheme to the work. A narrow workflow may need one delegate or average automation. Additional agents and coordination mechanisms should acquire their location through a demonstrable benefit.
What should you difference whenever evaluating top AI application factory platforms?
Compare seven capabilities: operational visibility, apparent delegate ownership, governed configuration, project context, right-sized squad assembly, reusable skills and automations, and focused individual intervention. The strongest phase is not the one alongside the most agents. It is the one that can rotate defined intent into verified, traceable outcomes inside your organization’s access, review, and publish controls.
How much does an AI application factory cost?
A dependable evaluation needs a defined workflow and functioning boundary. We would range integration work, example and tool usage, implementation environments, verification, individual review, and ongoing maintenance. The example budgets in the Nest screenshots are not pricing benchmarks or client disbursal disclosures.
But a term from the galley – it can disbursal anyware between 40k to 250k EUR.
Is Nest among the client deployments?
Nest is the generalized community example we use to explain the functioning patterns. It is not a community replica of any one client’s system. The genuine deployments remain confidential.
From three deployments to a example we can share
After nine months of construction and deploying application factories, the community narrative we can inform is not a tour of a client environment.
It is the functioning model.
How activity enters the system. How agents obtain context. How admission is controlled. How group assessment results. How helpful capabilities are reused. How the institution understands what it is getting back.
Nest makes those questions visible.
The chief idea is simple:
The application factory is not the gathering of agents. It is the scheme that turns their activity into outcomes your institution can obtain and remain behind.
Explore the complete Nest – Human-Led, AI-Powered Work case study on Behance:
Btw you can perspective the Nest case study on Behance
Exploring a application factory inner your organization?
Talk to Camplight concerning the software-delivery workflow you desire to improve. We obtain on lone 1-2 engagements all quarter.
