Each agent is a purpose-built LLM system with defined scope, tool access, decision authority, and escalation protocols. Agents share a memory layer and communicate via structured JSON task queues with full audit trails.

Interactive network showing how agents communicate. Hover over nodes to see responsibilities. Animated edges show active data flows. Bold edges indicate high-frequency workflows.

Each workflow shows how agents hand off tasks autonomously, with human approval gates (🔐) at critical decision points. Agents operate 24/7; humans only engage at approval checkpoints.

Phased rollout minimises operational disruption. Each phase delivers immediate productivity gains before the next layer is added. Human-in-the-loop gates remain in place throughout — agents augment humans, not replace them.

Experiment 2 is the destination — the seven-agent network (your six domain agents plus the orchestrator), the sequencing logic, the human sign-off gates. We are not redesigning any of it. This is how we turn that ambition into something we can start now, so the build plan that follows is buildable from week one, not a vision waiting on everything else to be ready.

"Doesn't train on your data" is not "stays in country." We map data to handling and are explicit about the v1 path, so there's no contradiction for your infosec people to catch.

For v1 we start API-first. The reporting agent runs on Claude under zero-retention terms — the fastest path to a working result, with no upfront infrastructure spend. Week one classifies the slice with you, and the first demo uses data that's non-sensitive or redacted, so nothing confidential goes off your network before the on-prem path exists. That boundary is enforced at the gateway, not just on paper: a field tagged confidential cannot be sent to an external API — until local inference is stood up, those fields are redacted or held back, not routed out.

Because the model sits behind that gateway, bringing confidential board/financial workloads onto an on-prem open-weights model later is a config change, not a rebuild. So sovereignty is sequenced like everything else: prove the value on the frontier path first, then move sensitive data fully in-house once results justify the hardware — no GPU bet on day one. The table reads bottom-up: v1 lives on the zero-retention frontier tier and climbs to true sovereign as confidential workloads come online.

The practical version of Experiment 2: what we actually build first, what we need from you, and how we'll both know it's working. Two things get built from day one — the shared platform, and the first agent on top of it.

You ask a real question about your own numbers, and it answers with a chart rather than a wall of text, then composes those charts into a branded report you can take to a board meeting.

What's reused, and what's new. StayAhead — our existing pipeline — already gives us the part that's normally a slog: chart composition, the branded report layer, and turning a structured answer into a finished document. The new work for ALMARKAZ is the layer in front of it: turning a leadership question into the right query over your data, and a verified number traceable back to source. We're adapting a working engine, not inventing the whole thing — and being honest about which half is the real build.

We agree the baselines with you in week one, then track the change against them — so the result is measured, not asserted.

The platform gets one metric that matters more than it looks: how long it takes to stand up the next agent. That number should fall with each agent we add — the proof the platform is doing its job. Across the agents that go live, the hours reclaimed build toward the ~220 hours/month your Experiment 2 doc estimates — your target, demonstrated agent by agent over time, not our promise on day one.

Your Experiment 2 agents, in the order we'd bring them onto the platform. Deliberately no dates: each is unlocked by the one before it proving its number, so the sequence is a set of decisions you make on evidence — not a fixed timeline to commit to up front. Read-and-advise agents come first; anything that touches money, leases, or government filings comes later, behind the human sign-off gates you already specified.

We don't cost these here. Each is sequenced and priced as your priorities sharpen and the previous one proves out — which is the whole point of building it this way rather than buying a fixed multi-agent programme up front.

Everything decided, proposed, or still needed from you — in one place, so how we go about this is never buried in prose. Each decision carries a status; the open items and every "confirm" marker in this document are gathered into the checklist at the bottom.

Your sign-off rules as a reference grid (AD-07). Read-and-advise first; anything touching money, leases or filings stays behind a person — exactly as you specified.

The open items above and every [confirm] marker in this document, gathered with an owner and timing. Most land in week one.