The agentic stack. Eight layers, one boundary.
A new lifecycle reuses everything below the domain agents.
Open any layer of the stack.
01
Signal intelligence
02
Model layer
03
Domain-specific agents
04
Execution Council
—
The boundary
A typed schema: nothing crosses it without a decision.
05
Routing layer
06
Execution layer
07
MCP server
08
Authentication layer
Thousands of narrow agents beat one wide one.
One master agent is proven in production, then copied per asset, venue or customer, all reporting to the same Council.
The swarm lab Illustrative
Running
- Proposal
- Settled
- Rejected
- Record
- 01 Signal
- 02 Predict
- 03 Decide
- 04 Approve ·
- 05 Execute
- 06 Settle
ProposalIn flight
Scaled down for your browser; a production swarm runs 5,000–50,000 agents.
Every action passes three gates.
Each gate leaves a record, designed so any action can be replayed.
- The boundary
Validate and risk-check 05 · Execute
Submit and confirm 05 · Execute
Reconcile and report 06 · Settle
- Records
Your agents see only what you grant.
Each customer runs on a dedicated, secure MCP server; any permission is revocable in one call.
Dedicated MCP server
Real-time data pipe
Permissions you can revoke
As autonomous as your mandate allows.
Switch modes at any time; the emergency halt works in all of them.
| Mode | Who approves | Latency |
|---|---|---|
| Fully automatic | Execution Council | milliseconds |
| Auto-with-approval | Operator or delegate | seconds to minutes |
| Manual | Operator | discretionary |
| Emergency halt | Any operator | immediate |
Built as a platform others deploy their own agents on.
The same eight layers, designed for third parties once trust and liquidity have been built in production.