One human director. A fleet of specialised agents. SOTA reasoning, governance, and delivery. Built on the same principles as Google DeepMind's Co-Scientist — in production, not a prototype.
The UNSW AI Institute is building world-class AI infrastructure across 300+ academics and 50+ research groups. The challenge is not access to AI models — it is orchestration: directing AI at complex, multi-stage problems with auditability, institutional memory, and accountability.
PCM (Pantheon Co-Scientist) addresses exactly this. It is a production-grade platform where a human director commands a fleet of specialist AI agents — each with a defined role, persistent memory, and structured output — through a governed process. It is already running in production across transport safety, financial trading, and platform engineering domains.
PCM is a multi-agent orchestration platform. The human director is the accountability anchor — the person who signs off, who is answerable to boards, funding bodies, and colleagues. The fleet of specialist agents does the reasoning, research, and analysis — at scale, in parallel, with institutional memory.
You. Sets the problem, reviews outputs, owns the decision. Agents work for you — not the other way around.
Coordinates the fleet, tracks dependencies, manages timelines. The agent that keeps everything moving.
Governance, strategy, risk assessment. The agent that asks: is this the right thing to do? Is this defensible?
Hard engineering, algorithms, data pipelines, implementation. The agent that builds the thing correctly.
Visualisation, presentation, user experience. The agent that makes the output understandable and compelling.
Market intelligence, external research, competitive positioning. The agent that challenges assumptions.
The GAIOS backend is production infrastructure, not a prototype. It runs on a tiered LLM architecture with persistent memory, canonical truth stores, and a governance layer.
Six minister agents, each potentially on a different LLM — producing genuinely diverse perspectives rather than groupthink from a single model.
Canonical documents, vector search (ChromaDB), and structured state graphs (SPINE) — institutional knowledge that survives across sessions.
All significant outputs pass through two independent adversarial passes via GPT-5.5 before ratification. Bad ideas die before they reach the director.
6+ months live uptime. 32/32 calibration tests passing. systemd services. Cloudflare CDN. Not a prototype.
In May 2026, Google DeepMind published Co-Scientist (Nature) — a multi-agent system on Gemini that generates, debates, and evolves hypotheses for hard scientific problems. Co-Scientist reproduced 10 years of antibiotic resistance research in 72 hours.
The Pantheon Co-Scientist platform (PCM) implements the same principles in production — with a critical addition: a human director in the loop.
Every project on PCM goes through a structured 9-step process. This is how complex problems get decomposed, stress-tested, and resolved — without defaulting to the first idea that arrives.
Sue Keay has 30+ years shaping Australian robotics and AI. She founded Robotics Australia Group, led Australia's robotics roadmaps, and is focused on translating AI research into practical applications. Her priority: AI infrastructure that can handle real-world complexity.
PCM can coordinate across UNSW's 50+ research groups — Engineering, Science, Business, Law, Medicine, Arts — without requiring every researcher to become an AI prompt engineer. The fleet handles the orchestration; researchers focus on domain expertise.
The 9-step inception protocol structures the path from idea → validated hypothesis → research plan. PCM's adversarial critique gates mean only vetted ideas progress — reducing noise and focusing resources on high-conviction directions.
Research projects often span years. PCM's canonical memory layer means the platform remembers institutional context, prior decisions, and previous findings — reducing redundant work and maintaining continuity across personnel changes.
PCM's multi-agent fleet architecture mirrors how complex robotic systems need coordinated subsystems. Applying PCM to autonomous systems research gives Sue a testbed for AI coordination at institutional scale.
Ian Gibson spent 30+ years in computer science and R&D management — including leading research at Canon's R&D lab and running Intersect Australia. His priority: demonstrating AI that delivers measurable industry and research outcomes.
This is not a research prototype. PCM has 6+ months production uptime, 32/32 calibration tests passing, systemd services, and Cloudflare CDN. The architecture is documented, auditable, and already delivering results in transport safety, financial trading, and platform engineering.
Ian's career is defined by taking research and getting it into products that ship at scale. PCM is built by an engineering team with the same orientation — the deliverable is not a report, it is a working platform that produces results.
GoldNet's engagement with UNSW is structured as a genuine partnership — Sabour Hosseini is the principal, and the PCM platform is offered for UNSW's use. This is not a vendor sale; it is a collaborative development relationship.
Co-Scientist demonstrated 10 years of antibiotic resistance research reproduced in 72 hours. PCM implements the same principles for any domain. For UNSW researchers, this means machine-speed hypothesis generation and stress-testing — accelerating the pace of publishable research.
The Institute spans 7 faculties. PCM's multi-agent fleet can parallel-coordinate across them — each faculty research agenda gets a specialist agent perspective, unified under the human director.
PCM's adversarial critique gates and ATHENA governance agent mean AI recommendations are stress-tested before reaching the director. For a research institute operating under public scrutiny, this accountability structure matters.
The tier cascade architecture routes to the most cost-effective model first, scaling up only when needed. For an institute managing 300+ academics, this is not a toy — it is infrastructure that handles volume.
Research commercialisation is a stated Institute objective. PCM can support this: structured inception of commercial ideas, adversarial vetting of go-to-market strategies, persistent institutional memory of partnership discussions.
The platform did not appear fully formed. It went through a structured evolution — mirroring the same 9-step inception process we propose for UNSW research projects.
Initial multi-seat architecture. Single-bot-per-seat model. No persistent memory. Functional but not production.
Migrated to self-hosted OpenClaw containers. Named ministers with distinct roles. First identity seeds authored. Fleet PM authority scoped.
Read Google DeepMind's Co-Scientist (Nature 2026-05-19). Mapped its seven specialised agents onto our six ministers. Identified five structural gaps. First proposal of 9-step inception protocol.
Integrated plan v1 + v2 (six load-bearing fixes). SOUL+context layer. SPINE dependency graphs. Two adversarial critique passes via GPT-5.5 PC bridge.
Minister-edge gateway. Preamble auto-construct. 32/32 calibration tests pass. Full observability layer operational.
PCS Heartbeat v2 spec critic-vetted. Automated state monitoring across all six minister agents. DRY_RUN soak active. System fully hardened for production multi-domain operation.
PCM offered to UNSW AI Institute as a collaborative development partnership. Live platform demo, institutional onboarding design, and first research project scoping.
Dr Sue Keay (Director, UNSW AI Institute) · Prof Ian Gibson (Deputy Dean)
Sabour Hosseini · GoldNet Group
Platform live: analysis.goldnetgroup.com.au · trading.goldnetgroup.com.au · vicrisk.goldnetgroup.com.au