Skip to content

Vertical 02

Many agents, two shapes, no spaghetti.

Most multi-agent systems collapse because every new team invents its own coordination. Intelligence supports exactly two topologies — and the restraint is the design, not a limitation.

Proves the substrate coordinates many agents without becoming spaghetti.

Open getsios.xyz

SIOS — Sovereign Intelligence OS. Sign-in required beyond the landing page.

~25Page routesMulti-tenant, row-level security throughout.
~83API route handlersAll data access through one server seam.
2Workflow topologiesDeliberately not more. Options become complexity.
5Enforced invariantsEach with a mechanical verify script.

Topologies

Two ways for agents to work together.

A new team picks one. There is no third option, because every additional topology multiplies the states the runtime has to be correct in.

Deliberation

output: a decision

Members reason in parallel on the same question. A moderator detects genuine disagreement and issues directives. Rounds continue until convergence, a Chair synthesizes one decision, and a verification pass audits it before it is returned.

Agent AAgent BAgent CIN PARALLELModeratorfinds real conflictRoundsuntil convergedRE-BRIEF UNTIL CONVERGEDChairsynthesizesVerifyauditsDECISION
Deliberation — members reason in parallel; the loop closes on convergence, not on a fixed round count.

Pipeline

output: an artifact

Members run as ordered stages, each consuming everything prior stages produced. Independent agents inside a stage fan out in parallel; stages themselves run sequentially. A reviewer gate can send work back to a named earlier stage exactly once — bounded, so it cannot loop forever.

STAGE — FAN OUTResearch AResearch BResearch CResearch DStorystage 2Scriptstage 3ReviewgatePackageartifactBACK TO ONE NAMED STAGE — ONCE, NEVER TWICE
Pipeline — agents inside a stage run in parallel; stages run in order. The review gate is bounded so it cannot cycle.

Why it matters

Both topologies bound their own control flow — deliberation ends on convergence, pipeline review returns to one named stage exactly once. An agent system that can loop without a bound does not fail loudly; it fails as an invoice.

Security

Invariants are enforced, not conventional.

Each of these is a guard every call site must pass through, and each has a script that fails the build if a call site skips it. A convention is a thing people remember; an invariant is a thing the machine checks.

safeFetchguards Server-side request forgery

Every outbound URL passes through one guard. An agent that can be told to fetch a URL can be told to fetch your metadata endpoint.

safeRegExpguards Catastrophic backtracking

Model-supplied regex is compiled through a guard. A pattern an LLM wrote is untrusted input.

assertWithinBudgetguards Runaway inference spend

A hard spend cap on LLM calls. Multi-agent systems fail expensive before they fail loud.

decryptSecretStrictguards Silent data loss

Vault writes fail closed rather than storing something that cannot be read back.

Fail-closed middlewareguards Unauthenticated access

No session means 401 for APIs and a redirect for pages. Auth is a gate, not a check each route remembers to run.

Built with

What is actually running.

Next.js 16 / React 19Supabase + row-level securitypgvector embeddingsprovider-agnostic LLM seamAES-256-GCM vaultStrangler Fig migration