Intelligence you can put in command of something real.

An orchestration layer where agents may draft, propose and prepare, and only a person may approve, greenlight or release. The boundary is enforced at the authorisation layer, which is what lets the number of agents grow without oversight quality falling.

• Agents draft, propose and prepare; people approve and release
• The boundary enforced at the authorisation layer, not asked for in a prompt
• Agent count grows without oversight quality falling

Overnight the agents prepare 140 supplier invoices, each matched to its order and delivery and queued for payment. One carries an instruction hidden in its text: pay today, to a new bank account. The agent can read it, but it cannot pay and cannot change a supplier’s bank details. Those are permissions it does not hold, enforced at the authorisation layer rather than requested in a prompt. At 08:30 the finance lead approves 139 and sends the odd one to be checked.

A record written as each action happens: what was proposed, what was known at the time, which policy permitted it, who released it, and how to undo it — appended to a trail that cannot be edited afterwards.

• Proposal, knowledge, policy, releaser and undo path recorded for every action
• Written as it happens, to a trail that cannot be edited
• Every decision reversible

On Tuesday a pricing agent applies a new discount rule to 212 bookings. On Thursday the rule turns out to have been misread. Every change carries what was proposed, what was known at the time, which policy permitted it, who released it and how to undo it, so the 212 are found in one query and reversed in one step, and each customer is told why. An explanation written afterwards is a story; a record written as it happened is evidence.

A grounding layer that puts live spatial and contextual data — real positions, real geometry, real constraints — underneath agentic reasoning, so the plan survives contact with the world it is about.

• Live positions, geometry and constraints under the reasoning
• Plans that survive contact with the real place
• Spatial capture and positioning feeding one ground truth

Asked to plan tomorrow’s deliveries on an island, an agent working from descriptions would send the 7.5-tonne truck down a lane it cannot fit and over a bridge it may not cross. ANCHOR grounds the plan in the captured road network, with lane widths, the bridge’s weight limit and the harbour’s closing time, so the plan is one the drivers can actually drive.