When ChatGPT 3.5 launched, I had the sinking feeling that we were going to rebuild everything every six months.
We more or less did.
Models didn't know our businesses. We built RAG.
They forgot. We built memory.
They couldn't finish anything reliably. We built agents, planners and orchestration.
They did strange things. We built evals and guardrails.
Then we built harnesses around all of it and sent very smart engineers into companies to make the whole contraption work.
Four years later, things had even begun to look suspiciously like architecture.
And then came Astra.
We may have put scaffolding on the asset register
The interesting thing about Astra isn't another benchmark going vertical.
ARC Prize put it into unfamiliar environments. Astra inferred the rules, created compact representations of the worlds it encountered and invented its own shorthand to reason about them. Give it a coding environment and it started writing parsers, state models, search algorithms and planners. For some games, it effectively built little software libraries for itself.
The model was building parts of its own harness.
There is an irony here.
Astra itself still benefits enormously from a good harness. ARC Prize got 62.7% with its standard setup and 99.9% with a provider-specific adapter.
So, no, the harness isn't dead.
Scaffolding is useful too. Nobody asks the builder to remove it while he is standing on the third floor.
The mistake is assuming it is part of the building.
Every predictable deficiency in frontier AI has produced an industry response. Eventually the workaround gets APIs, startups, consultants, conferences and a reassuringly capitalized category name.
At which point the model improves.
So I find myself asking a slightly uncomfortable question about almost every piece of the current AI stack:
Does this exist because reality requires it, or because the model is still stupid?
Planning?
Orchestration?
Memory?
Context management?
Ontologies?
Workflow decomposition?
Some will survive. Some will disappear into the model. Some will simply be generated on demand.
Today's scaffold is tomorrow's model capability.
Which brings us to the FDE
The Forward Deployed Engineer is having a moment.
TCS wants thousands of them. AI companies want them. Investors love businesses that have them. The FDE has become the Navy SEAL of enterprise AI. Drop one behind enemy lines with a laptop, API keys and access to the COO, and useful AI will eventually emerge.
There is a good reason.
An enterprise isn't a benchmark.
It is thirty years of software, five acquisitions, three ERPs, undocumented exceptions, employees who know things nobody bothered entering into Salesforce and a spreadsheet called FINAL_v7_USE_THIS_ONE.xlsx.
Someone has to understand all of that and make the AI useful.
Today, the FDE does.
But look at the job closely.
Understand an unfamiliar environment.
Discover how work is actually done.
Represent it.
Find the relevant tools.
Build missing ones.
Break work into steps.
Keep state.
Deal with exceptions.
These happen to be precisely the abilities moving into the model.
The FDE is solving a very real problem.
I am less convinced that it is a permanent one.
There is a harder problem hiding behind it
Imagine the model gets absurdly good.
It understands your ERP.
It watches the finance team for a week and understands why the numbers in the ERP aren't actually the numbers the CFO uses.
It knows the customer.
It knows what happened the last six times this customer asked for the same thing.
It can operate the software.
If necessary, it can write the missing software.
Wonderful.
Can it refund $20,000?
Can it change the price?
Can it submit the claim?
Can it reject the claim?
Can it sign the purchase order?
Can it reschedule the factory?
Can it promise delivery on Friday?
Can it fire the vendor?
Can it override an employee?
Ah.
We don't have an intelligence problem anymore.
We have an agency problem.
Knowing what to do and having the institutional right to do it are two entirely different things.
That difference gets more important as models get smarter, not less.
Agency Engineering
I think this becomes a discipline.
Agency Engineering is the engineering of how machine intelligence is allowed to participate in an enterprise.
What can it see?
What can it remember?
What state does it inherit?
What can it decide?
What can it do?
For how much money?
For how long?
When does it need another agent?
When does it need a human?
Who carries the can when it gets things wrong?
And, perhaps most importantly, what does the company learn after it acts?
We already have bits of this everywhere.
IT owns systems.
Security owns access.
Legal defines constraints.
Operations owns outcomes.
Compliance appears, as compliance reliably does, when everyone else has finished having fun.
Agency cuts across all of them.
Microsoft's Agent ID work is an early tell. Agents can have identities of their own. They can have scoped access and lifecycle controls. More interestingly, Microsoft requires agent identities to have a human sponsor responsible for their business purpose and access.
Read that again without the Microsoft terminology.
The machine has an identity. It has authority. And a human has to answer for it.
That isn't really an IAM feature.
It is the primitive beginning of an organizational model for non-human actors.
IT was built for systems. This is about actors.
Traditional enterprise software mostly sat there until somebody clicked something.
An agent doesn't.
It observes.
It reasons.
It chooses.
It acts.
And, increasingly, it figures out how to act.
Treating this entirely as an IT problem is a little like saying HR is responsible for employees because HR issued their laptops.
IT will be a large part of Agency Engineering.
It just won't be all of it.
Because the new unit of design is no longer only the system.
It is the actor.
This changes how I think about FDEs
The FDE approaches the enterprise from the outside:
Here is intelligence. Let me make your company understandable to it.
Agency Engineering starts inside:
Here is our company. How do we allow intelligence to act inside it?
That sounds like a small difference. I suspect it is an enormous one.
Because if Astra's trajectory continues, machines will get progressively better at understanding companies themselves.
The durable competence will then not be translating the enterprise for the machine.
It will be designing the enterprise so machines can safely acquire and exercise agency within it.
That competence has to compound inside the enterprise.
Otherwise something odd happens.
OpenAI understands your company.
Palantir understands your company.
The FDE understands your company.
The next FDE understands your company.
But your company never gets particularly good at deploying intelligence without them.
That cannot be the end state.
Build around what the model cannot swallow
It would be silly to conclude that RAG dies, FDEs disappear or every orchestration startup goes to zero because Astra had a good week.
That isn't the point.
The point is to be careful about declaring anything permanent when you are building immediately above an exponential.
There is one useful test.
Can greater intelligence make this layer disappear?
If yes, watch it carefully.
But there is another class of things intelligence cannot swallow.
Money.
Contracts.
Customers.
Permissions.
Regulation.
Authority.
Consequences.
Accountability.
These don't exist because the model is weak.
They exist because the enterprise is real.
Which is why the next important enterprise discipline may not be about squeezing more intelligence into the company.
We have become reasonably good at that.
It may be about giving intelligence agency without giving away the company.
We spent the last few years engineering intelligence into enterprises.
Now comes Agency Engineering.