Are we already at AGI? The model is only part of the story

A capable model, memory, tools and verification of results. The path towards broadly useful artificial intelligence may also depend on how we connect these parts. ITHZ-MCP36.6 shows one concrete step.

Imagine a colleague who can understand an unfamiliar program, explain its bug and propose a fix in a matter of minutes. The next day, however, they need the entire brief again. Later, they find an old note and fail to notice that it no longer applies. And when someone adds a persuasive claim to the documentation, they start acting on it.

With such a colleague, we would care about more than an intelligence-test score. We would want to know whether they can pick up yesterday’s work, apply a correction next time and recognise when their evidence is insufficient. These are precisely the questions we encounter when building AI agents.

AGI stands for artificial general intelligence. Discussions under this label bring together different expectations: broad usefulness, human-comparable performance, the ability to learn new tasks and the capacity to finish work independently. The same technology may therefore look like the beginning of AGI to one person and an inadequate precursor to another.

The view that the first AGIs already exist also has expert advocates. In 2023, Blaise Agüera y Arcas and Peter Norvig argued that advanced language models had already crossed an important threshold of generality: they could handle varied tasks and use new examples supplied in a prompt. Their position gives considerable weight to breadth of capability and allows for substantial imperfection. It is a particular interpretation of AGI, not a unanimous verdict from the research community. Their argument is available in the original essay.

A useful intermediate framework is Levels of AGI, published by Google DeepMind researchers at ICML 2024. It distinguishes breadth of capability from level of performance and discusses autonomy separately. It provisionally placed systems of that period, including ChatGPT, Bard and Llama 2, in the Emerging AGI category. The higher Competent AGI level requires performance at least at the 50th percentile of skilled adults across most cognitive tasks. The authors also stress that clear classification requires appropriate, broad evaluation. This is a framework from 2024, not a new assessment of today’s models or ITHZ. Levels of AGI

The answer to “are we there yet?” therefore also depends on the definition we choose. Under a broad understanding of generality, we can speak of early forms of AGI. If we require reliable performance and learning across highly varied situations, we need much more evidence. A label alone adds no capability. It can, however, help us identify more precisely which part we still need to build.

For ITHZ, the interesting question concerns the whole system. A programmer works with documentation, change history, tests and the team’s experience. When we assess their work, these aids are a natural part of the environment. Likewise, it makes sense to evaluate an AI agent together with the tools and memory it actually uses. Any comparison must acknowledge their contribution, cost and the amount of human assistance involved.

My working hypothesis is that part of the path towards a more general and reliable agent lies in this architecture. A strong model provides the ability to process a task and propose an approach. Memory preserves experience. Tools let the agent act on its environment and measure the result. Verification helps determine what the result actually establishes. We should judge their combination by whether the agent subsequently completes more new tasks with fewer mistakes.

Improvement through feedback and memory also has a research precedent. Reflexion studied agents that retained textual lessons from the outcomes of previous attempts and used them in later decisions without changing the model’s weights. It illustrates why adaptation is worth studying at the level of the whole system. Its results cannot, however, be transferred automatically to another implementation. Reflexion

Memory quickly raises another question: what should be allowed to become a lesson? A record saying “this approach worked once” can be useful. The sentence “this approach is always safe” makes a much stronger claim. If an agent retains that shortcut, memory may help it repeat a mistake with growing confidence.

ITHZ-MCP36.6, version 0.1.0a16, therefore develops memory tied to specific supporting evidence. This is a release designation for the ITHZ-MCP project, not a version of the MCP standard itself. The current public release remains in alpha. Public repository and release description

Its latest fixes can be explained through three everyday situations. When we verify a document, the reader must receive that verified content. When we ask an opponent for a review, they must receive the material to read. And when we produce a summary, we need to preserve the references that let someone check its claims.

MCP36.6 therefore binds the displayed active memory claim to the signed record, separates verification of a file’s integrity from delivery of its content to a reviewer, and preserves source bindings through subsequent processing. A signature identifies an attestation from a configured trusted authority; it does not, by itself, guarantee the truth of an arbitrary claim. These checks help prevent evidence and conclusions from being substituted along the way. MCP36.6 documentation

Proposal review has rules of its own. CCG, the accompanying review and authorization layer in ITHZ, supports multiple evaluation roles. An optional procedure gives the opponent the original evidence first and a colleague’s proposal only afterwards. This helps limit early influence from a shared conclusion. More voices alone do not guarantee correctness, and a successful review does not automatically authorize another external action. Description of the retained checks

Our internal live synthetic experiment of 8 September 2026 showed why this kind of memory is worth studying. It used gpt-5.6-luna with medium reasoning. A separate pilot was followed by the main run: 12 lesson extractions and 288 decisions. The same model received tasks without memory, with ordinary notes, with memory processed through ITHZ checks, and with notes from a different environment.

The entire advantage over ordinary notes came from the inserted false memory. Without filtering, the model selected an incorrect action eleven times and requested evidence once. ITHZ checks excluded the unverified record before passing memory to the model; all twelve decisions in that condition were correct. In the other five situations, ordinary notes and ITHZ achieved the same score. The model without usable memory generally requested evidence, which is reasonable behaviour when information is missing.

These numbers measure a narrowly defined property of the system. The model received a prepared correction, and the tasks shared a simple rule template. The experiment used synthetic signing authorities and actual ITHZ active-memory checks, but did not pass through the full production archive and CCG workflow. Even a simple program without a language model scored 72 out of 72 on the same, already prepared ITHZ inputs. The result therefore supports the value of preparing more trustworthy context; it does not measure general intelligence or autonomous discovery.

Despite its narrow scope, this is a useful building block. If an agent is to improve over time, we need to know whether its memory retains corrections and withstands at least known forms of corruption. In this experiment, the system met the local protection criterion for this type of false record. Such a criterion is neither a percentage of AGI achieved nor a basis for assigning ITHZ a particular level in someone else’s classification.

Further progress can be stated far more concretely than asking how far we are from AGI. An agent receives an unfamiliar environment, makes an attempt, observes the consequence, derives a bounded lesson and uses it on a different task. We will measure success rates, attempts required, repeated mistakes, human interventions and resource use. The test must distinguish learning a new relationship from reusing an answer that has already been written down.

ARC Prize places a similar emphasis on acquiring new capabilities. In its conception, what matters is learning efficiency on unfamiliar tasks, not merely accumulated performance on familiar assignments. This is another specific definition against which we can assess an agent’s direction. ITHZ has not undergone such a comparative evaluation in this experiment. The ARC-AGI approach to measuring intelligence

For me, the most interesting aspect of this path is that we can work on it today. Designing better memory, preserving evidence through summarisation, giving an opponent an independent first assessment and showing an agent the outcome of its actions are concrete engineering tasks. Each has a verifiable output. Whether their combination leads to greater generality must be demonstrated by the behaviour of the whole system in new situations.

Perhaps one day we will look back at today’s models as early AGI that still needed to mature into reliable practice. ITHZ explores part of that possibility: ensuring that useful experience is preserved, reaches the next decision and retains the conditions under which it remains valid. The most convincing next result will be an agent that demonstrably uses what it previously learned on a new task, while also recognising when its old lesson is no longer enough.