The control plane.
Sits above the applications, not inside one. It holds the operating model (what good looks like, the rules, the gates), the skills that run each workflow, the run log and the conversation.
Business systems are moving from applications that record and route to agentic surfaces that reason and act. That is a paradigm shift, not a feature, and we are driving it. This capability is how: the ability to architect, build, integrate and run complex agentic solutions for many clients at once, as managed solutions. The paradigm is on its own page; this page is the engineering.
The applications of record stay where they are. Everything else is what we build.
Sits above the applications, not inside one. It holds the operating model (what good looks like, the rules, the gates), the skills that run each workflow, the run log and the conversation.
A custom set of screens over the client’s applications, every screen carrying a conversation pane that knows the page, so people talk or click on the same screen. The same brain answers in the chat tools they already use.
An agent for each job: it reads across the systems, drafts the next step and runs the obvious ones. Anything that costs money, carries risk or needs taste waits for a person’s one-digit answer.
The scheduled runs (nightly reads, daily observe and diagnose, the weekly slate, the monthly review) and the event-driven ones (a lead arrives, a reply lands, a threshold is crossed), each with a scope, a gate and a failure behaviour.
One performance table across every source, and one row per action in the run log: who asked, from which surface, who approved, what the API returned, what it cost. The audit trail and the source the reports read.
Gateway functions to the client’s own platforms through their APIs, credentials in an encrypted store no person or prompt can read, every write read back. Platform agnostic: any vendor in a category whose APIs carry the use case.
None of them is exotic. Together they are what makes an agent safe enough to sell and cheap enough to run.
The custom screens and, on every one, the conversation pane that sends the page context with each message and refreshes when the plane reports a change.
The operating model, the skills that run each workflow, the gates (thresholds and approvers per action), the conversation handling for every surface. Credentials live here, encrypted.
The client’s CRM, marketing, finance, help desk, inventory, analytics and the systems around them, called by the plane’s gateway functions; every write read back.
One row per action: who asked, from which surface and screen, the gate and who passed it, the call and its read-back, the cost.
The verbs are the same in every function we have built for. The example is our own marketing, because it runs today and you can watch it.
Themes scored against search demand, competitors and the pipeline every month; programs with a hypothesis, a window and a kill rule.
The week’s slate, the budget and the next move, read overnight from the numbers and posted with its reasons. Three decisions, one digit each.
Posts, videos, emails, social copy and ads made to the plan, each to its own gate, each with its cover, its tags and its cost on the row.
To the CMS, the video channel, the ad account and the email list on schedule, never before the page it links to answers, every URL checked in the same run.
One performance table from ads, analytics, search, CRM and email; qualified leads sent back to the ad platform; the all-in cost landed on every piece.
Measured on the solution we run for ourselves, AI Marketing OS, by industry-typical task times for the work as specified. These are the numbers we show a client, with the method.
The first twenty days of the build (1 to 20 September 2026) delivered 4,442 hours of work by conventional task times: 27.8 person-months of a conventional build, in 20 calendar days. That is 222 hours of output a day, 9.3 times one person working around the clock. A conventional team of the same size would still be in discovery.
As designed, the routines deliver about 1,380 hours of work a month for 43.5 hours of a person’s supervision: 8.6 full-time jobs replaced, and the management those jobs needed cut by 86 percent, from 2.27 to 0.31 full-time equivalents. The tooling costs 400 dollars a month before media.
| Function | Hours a month |
|---|---|
| Blog, twenty posts a week | 325 |
| Video, ten videos a week | 624 |
| Search health and crawl, weekly | 200 |
| Measurement, daily | 75 |
| Paid search, daily | 38 |
| Newsletter and weekly summary | 37 |
| Programs and themes | 24 |
| Influence, social, self-check, replies | 59 |
Each row is units a month times industry-typical hours per unit, never a measurement of our staff. Published trials of AI tools on knowledge work measure 0.84 to 1.67 times one person’s output; a governed system of agents is a different thing from a tool in a person’s hands, which is the point.
The labour that goes is the coordination, the re-keying and the management of both. The deciding stays with a person.
Building one agentic solution is engineering. Running them for many clients, safely, is the capability.
Your records, your credentials in your encrypted store, your gates and your approvers, your run log. Nothing shared with another client’s instance.
One codebase and one set of versioned skills run every instance. A fix or an improvement reaches every client as a release, not as a project.
Every skill ships with its tests; a release checks the production surface before and after the change and fails on any lost field. Nothing reaches a client untested.
Connections and routines are tested daily. A regression is restored to the code of record and read back in the same run; a disconnected connector is reconnected. You hear about it as a line in a report, not as an outage.
A system runs what it is given. A senior person reviews the plan, the themes, the targets and the results with you, so the inputs stay sound and the outputs stay worth paying for.
The software is licensed, which is the subscription, and sold with managed services, billed as a monthly managed services fee, and advisory services. The estimate for setting it up on your platforms is guaranteed.
The three eras of business systems and the six solutions.
The engineering practice behind the agents, the skins and the adapters.
How the running of a solution is staffed and measured.
Workflow automation moves a record from one step to the next along a fixed path. An agentic solution reads across several systems, drafts the next step, runs the obvious ones and stops at a gate for anything that costs money, carries risk or needs judgment. It is built from six parts: a control plane, one set of screens, agents with gates, routines, a data and audit layer, and an integration layer over the client's existing applications.
Every action that spends money or carries risk waits for a person's one-digit answer, and spending caps are enforced in code below the gate. A gate raised on a screen can be answered from a phone, and the answer lands on the same record. Each run writes one row to a ledger: who asked, from which surface, who approved, what the API returned and what it cost. The deciding stays with a person; the coordination and re-keying go.
Any vendor in a category whose APIs carry the use case. The control plane never calls a vendor directly; gateway functions own each platform's API, so the CRM, marketing, finance, help desk and inventory systems can be whatever the client already runs. Credentials sit in an encrypted store that no person or prompt can read, every write is read back and compared, and a read-only probe is never treated as proof of a write path.
Measured on the solution CodeStringers runs for its own marketing, the routines deliver about 1,380 hours of work a month for 43.5 hours of a person's supervision, which is 8.6 full-time jobs by industry-typical task times, with the management those jobs needed cut by 86 percent. The tooling costs 400 dollars a month before media. The figures are units a month times typical hours per unit, not a measurement of staff.
The solution is licensed, which is the subscription, and sold with managed services, billed as a monthly managed services fee, and advisory services. Each client gets its own instance with its own records, credentials, gates, approvers and run log, while one shared codebase and set of versioned skills run every instance. Connections are tested daily and regressions restored automatically. The estimate for setting it up on the client's platforms is guaranteed.
The first twenty days of the CodeStringers build, 1 to 20 September 2026, delivered 4,442 hours of work by conventional task times, which is 27.8 person-months of a conventional build in 20 calendar days. A client build starts with discovery, where the operating model, the gates and the approvers are written down, and the first routines go live on the client's own platforms rather than on a copy.
We take on both today, and we will say which one you are buying.
"Something isn't working." "We're outgrowing our tools." "We're afraid to make the wrong move." No-Risk Discovery is a short, practical conversation that gets you clarity before you commit to anything big. We'll tell you if we're a fit. If we're not, we'll tell you that too.