Models
Reasoning and generation, selected for the task.
BEYOND THE SINGLE MODEL
Intelligence
is a system.
Not a waiting room for the next foundation model.
A world of personal agents. Connected through purpose, memory and tools. Orchestrated into teams that can do more together—and improve through evidence.
COORDINATEEVALUATELEARNREPEAT
COLLECTIVE∞POSSIBILITY
PERSONAL AGENTSSHARED CONTEXTHUMAN JUDGEMENTSYSTEMS RESEARCH / CONCEPT VISUALISATION
THE FUTURE IS A RESEARCH QUESTION.THE FIRST USEFUL WORKFLOW IS A DESIGN DECISION.
01 / THE COLLECTIVE INTELLIGENCE THESIS
More capable.
By design.
AGI has no universally agreed operational definition. Our interest is the system around a model: how specialised capabilities can be composed, checked and made useful across more kinds of work.
What if the next leap is not just a bigger mind—
but a better way for many minds to work?
Tools
Explicitly permitted ways to observe and act.
Memory
Relevant context with ownership and provenance.
Orchestration
Roles, handoffs, boundaries and stopping rules.
Evaluation
Evidence that a change makes the system better.
02 / THE COLLECTIVE LAB
One goal.
Many possible architectures.
Explore how team size, coordination and shared context change a system’s shape. More agents create more relationships to manage—not an automatic guarantee of better results.
ORCHESTRATOR + SPECIALISTSA clear owner.
A clear owner.
Specialised contributors.
An orchestrator routes work to specialist roles and assembles a result. It needs a defined task contract, tool permissions and a way to handle failed handoffs.
This is a topology illustration, not a benchmark, live agent service or performance prediction.
8 illustrated agent rolesMEMORY / SCOPED & SHARED
Composition Assign bounded roles to the work, not just more copies of the same prompt.
Coordination Choose who delegates, who reviews and who can stop the workflow.
Context Share what is useful while keeping client, personal and project boundaries intact.
03 / PRACTICAL INTELLIGENCE
Start with work.
Not a grand claim.
Useful business applications do not need to wait for general intelligence. They need a specific goal, permitted information, a reviewable output and an accountable owner.
RESEARCH & DECISIONS
A clearer brief.
A better-informed decision.
Coordinate research, synthesis and review around approved source material. Produce a source-linked comparison with assumptions and unanswered questions intact.
- Research role: collect permitted evidence.
- Analyst role: compare against agreed criteria.
- Reviewer role: challenge unsupported conclusions.
- Human owner: decide what to do next.
ILLUSTRATIVE DELIVERABLE
Decision brief
Sources → comparison → uncertainty → review
No decision is made or external action performed by this example.
04 / THE IMPROVEMENT LOOP
Learn from the work.
Earn the next version.
Self-improvement should mean more than generating another answer. Capture failures, propose a change and test whether it holds up beyond the examples used to create it.
Prompts, workflow rules, tool selection and memory policies can all evolve. Candidate versions should earn promotion through evaluation, with a person responsible for the release decision.
Read the improvement-loop guide →- 01Observe a failureREADY
Keep the case and its expected outcome.
- 02Propose a candidateWAITING
Change a bounded part of the system.
- 03Evaluate beyond the exampleWAITING
Include held-out cases and regressions.
- 04Human release decisionWAITING
Review evidence before promotion.
- 05Monitor the new versionWAITING
Keep a rollback path and compare outcomes.
An improvement is a hypothesis until it survives a test.
Illustrates an engineering process. It does not retrain a model or demonstrate autonomous recursive self-improvement.
05 / AN API-FIRST STARTING POINT
Build with what exists.
Design for what comes next.
Model APIs, retrieval, conventional software and controlled tool use are available now. The opportunity is in composing them into a useful system—not claiming every component must be invented again.
YOUR SYSTEM / YOUR BOUNDARIESILLUSTRATIVE ARCHITECTURE
// A task contract, before a tool call.
const mission = {
goal: "Prepare a source-linked brief",
sources: approvedDocuments,
tools: ["retrieve.read"],
review: "human-before-external-action",
evaluateAgainst: heldOutCases
};
// Adapt your chosen provider API.
// Keep secrets and execution server-side.The model is an interface.
The system is the product.
Choose a model API for a bounded role. Keep the task state in your application. Validate outputs, separate read tools from write tools, and measure the workflow against the outcome you actually need.
- Provider adapters rather than a hard-coded dependency.
- Explicit schemas and tool permission boundaries.
- Evaluation and observability from the first useful version.
Explore a browser-only request simulator, then the integration patterns behind real systems.
06 / INTELLIGENCE NEEDS AN EDUCATION
Teach the system.
Educate the people.
A capable team understands its tools, evaluates its work and knows when to ask for help. The same discipline belongs in the systems we build.
What would count
as general intelligence?
Separate broad ambition from narrow evidence. Define the tasks and conditions a claim must survive.
Explore the question →Make the learning
loop inspectable.
Version changes, hold out test cases and watch for regressions before increasing authority.
Build a better test →Train the person
behind the agent.
Write clearer missions, recognise uncertainty and review consequential actions with intent.
Start the practical guide →A LONG-TERM VISION. A USEFUL NEXT STEP.
The frontier is open.
Bring something worth building.
We welcome businesses with real problems, researchers with testable ideas, and partners or investors interested in the infrastructure of collective intelligence.
07 / START A CONVERSATION
Ambition is welcome.
Specifics are useful.
Tell us the workflow, the people and the constraints. Start a conversation about a practical application, agent architecture, API integration, research or a longer-term partnership.
An enquiry does not create a subscription, investment commitment or deployed agent. Please do not include credentials, confidential documents or sensitive personal data.
CLEAR ANSWERS / NO SHORTCUTS
Keep the possibility.
Keep the precision.
Has agi.pa achieved artificial general intelligence?
No. This site explores a systems thesis: specialised agents, tools, memory and evaluation may enable broader capabilities. It does not claim demonstrated AGI or proven autonomous recursive self-improvement.
Can we build useful systems with APIs available today?
Yes. Current model APIs and conventional software can support bounded workflows, tool use and coordinated agents. Every integration still needs permission boundaries, evaluation and operational ownership.
What does self-learning mean in this context?
The improvement loop describes versioned changes to prompts, workflows, memory rules and tools, tested against representative and held-out cases before human-approved promotion. It is not a claim that model weights retrain themselves.
Are the interactive explainers connected to real agents?
No. They are deterministic browser illustrations. They do not call models, access files, provision cloud resources or execute business actions.