Where do we go from here
Today at Friends House, Sir Demis spoke about the path to AGI.
Specifically, what is necessary.
His answer:
- Scaling laws will march on
As we unlock orders of magnitude more compute, the size of the models we can pretrain — and then RL-train and distill — will scale with it.
- Architectural breakthroughs
Model intelligence is still jagged, and learning is still mostly up-front rather than continual.
Sir Demis said it is not yet clear which of the two gets us all the way there, so DeepMind are pursuing both.
I think there is a third pillar: substrate.
No matter how far scaling laws, world models, or continual learning progress, the thing we call a model is, in principle, simple.
[ 0.5 -0.8 0.1] [0.8] [ 0.1] [ 0.13] [0.13]
[-0.2 0.7 0.4] * [0.5] + [-0.2] = [ 0.11] -ReLU-> [0.11]
[ 0.3 0.1 -0.9] [0.3] [-0.5] [-0.48] [0.00]
W x b z out Sequences of matrix multiplications and other simple mathematical operations, scaled to a degree that the human mind simply cannot comprehend — 500 septillion floating point operations to train, and then countless more to run inference.
A checkpoint may encode enormous capability, but it is still just a set of floating point numbers. By itself it has no continuity. It does not wake up tomorrow. It does not resume a task. It does not notice that the world has changed.
Then come agents: through some mix of semi-identityA prompt, name, and memory give the agent something approaching 'identity', but they can be paused, cloned, branched, merged, in a way that human concepts of 'identity' cannot. and memory, tools, and sensory apparatus, we give those numbers a way to persist long enough to do work.
That surrounding system is the substrate.
By substrate I mean the layer around the model that gives it continuity: state, permissions, feedback loops, and access to the world.
If scaling and architecture give us intelligence, substrate gives that intelligence somewhere to stand.
A model checkpoint is a brain in a jar.
Useful, but inert.
It does not perceive the world, does not remember yesterday, cannot decide to investigate something and come back to it later.
Those are not automatic by-products of intelligence. They are properties of the system around the model.
You can scale a model dramatically and still get something trapped in the shape of a single interaction unless the substrate lets it hold state, take actions, observe consequences, and iterate.
So: scaling determines how much capability is latent in the model. Architecture determines how that capability is shaped. Substrate determines whether it can persist, act, and compound over time.
This is why better tooling matters so much. Tooling is what lets a model do things — carry context across sessions, invoke APIs, run code, observe failures, and try again. The model supplies the intelligence. The substrate turns that intelligence into a process.
2025 was supposed to be the year of the agent. 2026 is now the year of the agent. Agent 'sandboxes' are raising serious amounts of capital to wrap Firecracker. New libraries for tool discovery, tool execution, tool composition, memory, compaction, and orchestration appear every day.
Clearly: software is not dead. What is the substrate if not software?
'Tooling overhang' is what happens when a technology advances faster than the tools that make it usable. The capability exists, but it is not yet available to the average person.
Post-2022 AI has created a tooling overhang on a scale we have not really seen before.
Right now the focus is on sandboxes. That makes sense, but a sandbox is only one primitive in what a real substrate will become.
If agents are to become genuinely autonomous, they will need a substrate that lets them build, deploy, and operate software.
Build.
Code generation is the least interesting part of building systems. The hard part is systems integration: verifying behaviour against third party systems, stitching components together, and closing the loop when reality diverges from the spec.
This will require tooling.
Deploy.
Shipping to production is not actually in a terrible place. Post-2022 AI already makes a lot of infrastructure work easier. But deployment still needs permissions, policy, rollback, staged rollout, and guard-rails. Humans need them; agents will too.
This will require tooling.
Operate.
This is where the gap is widest. Many teams are trying to solve human software operations pain via 'AI SREs' and the like. Useful, yes. Sufficient for autonomous engineering systems, no.
Agents will need first-class access to telemetry, diagnosis, change management, and long-horizon feedback loops.
This will require tooling.
So, where do we go from here?
In the next ~12 months we will see many companies try to expand from 'sandboxes' — which are quickly becoming commodities — into something closer to a full agent substrate.
Many have tried to build this before. The usual mistake is assuming an 'agent' is a bag of prompts and a few API calls.
It is not. The agents that matter will be long-lived systems with memory, tools, permissions, telemetry, and the ability to change themselves safely.
That is where software goes next.