We have enjoyed our time working on this — the Hidden Guild — website. But the technical edge of AI is already moving in a new direction. Which means it may be time for the Anti-Dave to reveal a bit more of his thinking about where the Future will be landing.
Because, dear Reader, that is what the Future does.
You step up to the tee, take your best swing and sometimes land in the middle of the fairway. Just as often, though, the ball disappears into the knee-high rough and you spend the next half-hour looking for something you were certain you understood five seconds after impact.
AI is now heading into that rough. Not because it is failing, but because it is becoming useful enough that the next problem is no longer simply intelligence. The next problem is organizing, packaging, testing and maintaining what the intelligence has been asked to do.
Two New Domains
For the “uninwebbinated,” a domain name is simply an address on the internet. But if you have read the ancient marketing classic Positioning by Al Ries and Jack Trout, you will appreciate the larger idea: A project’s name should tell people what shelf in the mind it belongs on.
With that in mind, we have registered two new domains.
Taskerware.com
The idea behind Taskerware is simple: AI is quickly moving beyond the point where every tiny procedural detail must be explained from scratch during every conversation.
AI is growing up. A decent amount of fault tolerance is already being built into these systems. Interactions are becoming smoother across platforms, including the smaller but increasingly functional 2-billion- to 8-billion-parameter models that can run on processor-bound personal computers without massive memory bandwidth.
Not all systems have embraced what we have long called the Shared Framework Experience: the collection of assumptions, definitions, methods and prior decisions that allows a human and machine to work together without starting over every morning. But the direction is obvious.
The intelligence is becoming plentiful. What remains scarce is a reliable description of the job.
That is the Taskerware layer.
A Taskerware package describes what must be accomplished. It contains the objective, required knowledge, available tools, permissions, sequence, standards, exception rules, completion tests and expected deliverable.
It is not merely a prompt. It is the durable specification of a task.
AI-OpCode.com
The companion domain is AI-OpCode.com.
Our working distinction is:
Taskerware packages how a job is done. AI OpCode tells machines how to execute, test and maintain it.
That is better territory than building “another agent builder.” Microsoft, OpenAI, UiPath and a growing herd of extremely well-funded companies can own the general-purpose execution machinery.
The more defensible layer resides in the human-derived task specification: the objective, knowledge sources, tools, permissions, sequence, exception rules, completion evidence and maintenance history.
Taskerware sits closer to workflow analysis and deliverable design. AI OpCode sits closer to the executable calls that make the workflow happen.
A Taskerware package might say:
Produce the Tuesday market report using these sources, these calculations, these chart standards and these editorial rules. Deliver a WordPress-ready article, a subscriber PDF and a short public summary.
The AI OpCode would handle the procedural layer:
On Tuesday at 4:30 a.m., load Task Package 117, collect the current inputs, run the specified analysis, validate the result, distribute it to List PN-Tuesday, update the website and open a feedback page.
One defines the job. The other operates it.
Plugging Past Plugins
The concept of the software “plugin” is about to undergo a remarkable change—one likely to affect server-farm operations, software maintenance and future demand for backend computing.
The walk-through matters because, at the highest level, no one has yet beaten a path to my door for the idea of a process-broker layer sitting above large shared computing systems.
But something like it will have to arrive.
Traditional plugins have usually been narrow additions to a larger program. The VST world in digital audio is a good example: One plugin provides an equalizer, another adds compression and another simulates a particular amplifier or room.
AI plugins will not remain that singular in focus.
To keep computing near the highest-performance and lowest-cost regions, plugins will become progressively more granular. One may hold instructions. Another may expose a data source. Another may contain a tool. Another may enforce a permission boundary, validate an output or format the finished product.
The intelligence will assemble what it needs for the job.
This is precisely where Taskerware sees opportunity jumping up and down, waving both arms.
Taskerware owns the “What is the deliverable?” layer. That includes not only the intellectual objective, but also mundane—and absolutely necessary—details such as file type, paper size, page orientation, naming conventions, citation requirements, approval points and distribution rules.
AI OpCode occupies the procedural layer: which task package to load, what tools to invoke, what order to follow, when to ask for human approval and what evidence must be recorded before the job can be declared complete.
The agent is not the product.
The completed, repeatable and maintainable job is the product.
Where Does This Place the Future?
We will not know until we get there, but several near-term product decisions are already becoming clear.
A Taskerware item should be a versioned package. It should contain a manifest, task instructions, source materials, required tools, permissions, acceptance tests and a change log.
The specification should also be separated from its execution engine. The same task ought to be runnable through OpenAI, Microsoft, a local model or even a deterministic conventional program. Model choice belongs in the deployment configuration—not in the permanent definition of the task.
Semantic versioning also makes sense. Correcting a prompt or documentation error might produce a patch release. Changing workflow behavior would justify a minor release. Breaking compatibility with existing inputs, outputs, permissions or tools would require a major version.
Migration and regression testing must be built in early. Each serious task package should include representative cases, including ordinary jobs, difficult jobs, known failures and exception paths. A new model, tool or connector should not become the “current” version until it passes those tests.
Maintenance must also become an explicit part of the product. The initial task build is one thing. Keeping it working through model retirements, API changes, connector failures, security discoveries and changing business rules is another.
The likely pricing model is therefore hybrid: a predictable subscription for maintaining the Taskerware package, with unusually expensive executions metered separately.
The opening is not:
We have smarter agents.
The opening is:
We preserve working knowledge as a portable, testable and upgradeable asset—even when the underlying AI changes or disappears.
What the Rest of the Field Is Thinking
The Anti-Dave has already sorted through several recent developments, and they point in precisely this direction.
OpenAI Is Moving from Custom GPTs to Plugins
OpenAI plans to retire custom GPTs on December 11, 2026, with a possible Enterprise deferral to February 11, 2027. Creation of new custom GPTs is scheduled to end October 26.
Their replacement is the plugin: a package that can combine reusable instructions, reference files, skills and connected applications.
Importantly, migration does not automatically preserve sharing permissions, and custom actions may need to be rebuilt. OpenAI advises creators to save representative prompts and regression-test the replacement because the migrated system may behave differently. OpenAI’s migration FAQ explains the transition.
That is not merely a product-name change.
The disposable custom chatbot is being replaced by an installable and maintainable software package. Instructions alone are no longer the whole product. Dependencies, permissions, tests, distribution and upgrade behavior now matter.
There is also a hard retirement date, so this is not speculative AI marketing. It is a lifecycle event.
Microsoft Has Split Casual AI from Agentic Work
Microsoft’s September 25 Copilot redesign divides the market economically.
Everyday questions, summaries, drafts and Office assistance remain under a fixed user-subscription model. Cowork, Code, Autopilot, frontier models and long-running agentic work move toward usage-based billing.
Microsoft is also introducing spending policies, credit-request approvals, model restrictions, usage histories and reporting intended to connect agent spending with business outcomes. Its Managed Runtime provides a governed environment for code, apps and workflows, while a unified plugin catalog is intended to let developers publish a capability across multiple Copilot surfaces. Microsoft describes the new Copilot structure here.
The meaningful shift is commercial. One of the world’s largest software vendors is validating a hybrid model: fixed subscriptions for ordinary assistance and metered billing for delegated machine work.
The hype boundary remains visible. Some of the new capabilities are only beginning limited rollout, while Autopilot remains in private preview. The pricing direction is real. The persistent digital employee is not yet a universal, broadly shipped reality.
UiPath Made the Workflow Specification an Asset
UiPath’s September 23 introduction of Cartographer may be the clearest confirmation of where the edge is moving.
Cartographer builds what UiPath calls a Map of Work: a tenant-owned, governed and versioned definition of how an enterprise process actually operates. It can collect knowledge from documents, systems, recordings and employees. Each asserted fact carries its source and verifier, while conflicting instructions are surfaced for human resolution.
Cartographer, Delegate and UiPath for Coding Agents are generally available. UiPath has also divided its Maestro product into Orchestrate for long-running, stateful business processes and Automate for large volumes of shorter tasks.
Most importantly, an approved Map of Work becomes the specification from which workflows, agents and tests can be built. Evaluations can then gate deployment. A planned Decision Ledger would capture production exceptions and turn them into proposed changes, subject to accountable human approval. UiPath’s FUSION announcement provides the details.
The market is finally treating workflow knowledge as something that must be sourced, versioned, approved, tested and maintained independently of whichever model happens to execute it.
That is the heart of Taskerware.
The Edge Is Moving
…and this means the frontier will follow.
There is a great deal of build-out left in “the Future.” It is possible that today’s apparent overbuilding of AI infrastructure will eventually return as a gift in the form of much lower per-token costs on distant backend compute systems.
But we would not hitch our wagons to that promise.
We are still waiting for atomic energy to make electricity so cheap that it will not be worth metering.
AI may follow the same path. Intelligence will become cheaper, but metering, permissions, maintenance and ownership will remain.
Which means the real edge may not belong to whoever owns the smartest model this quarter.
It may belong to whoever best captures how useful work is performed—and can keep that knowledge operating after the models, vendors and fashionable names have all changed.
Oh, and if you want either of those domain names? Everything is for sale; this is America after all. And the Anti-Dave needs more time in the shop and less playing in the future.
Because eventually it becomes the Past Ure.
~Anti-Dave