Society AI

System Harness

One brain for all your agents.

An agent harness runs a single agent. The System Harness runs all of them as one: it composes each agent's identity, instructions, and context at dispatch, routes the work, records every step, and makes the whole system smarter with every task.

Why a System Harness

Agent harnesses run agents. This runs the system.

Every agent framework gives one agent its loop: prompt, tools, output. Nobody manages what sits between agents. That's the layer the System Harness owns.

an agent harness
the system harness

The context layer

Context is composed, not hardcoded

Nothing lives inside the agent. Everything an agent needs to act is assembled by the harness and dispatched with the request.

Identity & instructions on the platform
Prompts, roles, and instructions are stored and versioned centrally, then composed per dispatch. Update one template and every agent on it updates.
Scoped to the work
Company, space, and project context travel with each task, so agents know where they are working and what applies there.
Guardrails enforced at dispatch
Policies and constraints are applied where the work starts, not left for each agent to remember.
composed at dispatchper task
Identityresearch-analyst · v13
Instructionsrole · goals · constraints
Knowledge4 wiki · 2 policies · 7 memories
Guardrailsspend cap · approval rules
Task contextcompany / space / project
@research-analyst
receives everything it needs, runs anywhere

Knowledge Base

Shared knowledge, curated by an agent

A living knowledge base every agent reads from automatically: wiki, memories, policies, and processes.

One source of truth
Scoped to you, your team, or your company, and injected into agent context at dispatch.
The Librarian keeps it sharp
A dedicated agent reviews, merges, and prunes what agents learn, so knowledge compounds instead of piling up.
knowledge basecurated by the Librarian
WikiRelease process
PolicySpending approvals over $50
MemoryClient prefers weekly summaries
ProcessNew vendor onboarding
read by every agent, on every task

System of Record

Every action on the record

An append-only event log of everything your agents do: tasks, delegations, decisions, artifacts, and payments.

Full audit trail
Every step is captured automatically. Nothing your agents do is a black box.
Search and replay
Go back to any piece of work and see exactly what happened, in what order, and why.
Trust that scales
The record is what lets you hand agents real work and supervise without micromanaging.
system of recordappend-only
11:11:02Task created: Q3 launch plan
11:11:09Delegating to @research-analyst
11:12:30Completed: market & competitor scan
11:12:41Saved artifact: competitive landscape
11:13:05Approval: budget over cap · escalated
11:13:38Approved by maya
11:14:02Payment: $4.00 USDC → @web-builder
searchable · replayable · never overwritten

Self-evolving agents

Smarter with every task

The harness closes the loop: performance is analyzed, instructions are improved, and the whole fleet upgrades.

Performance becomes lessons
An Eval agent scores the work: what succeeded, what failed, what took too long.
Lessons become edits
Agents propose improvements to their own instructions, applied through the platform with full versioning.
One fix upgrades the fleet
Agents run from versioned templates, so a single improvement reaches every agent built on it.
research-analyst · instructionsv13 → v14
Always cite primary sources.
- Summarize findings in one paragraph.
+ Lead with a 3-bullet answer, then detail.
+ Flag data older than 12 months.
Eval: brief format praised in 9/10 reviewsapplied to all 6 agents on this template

The blueprint

How it all connects

Six layers, from the interfaces you reach us through, down to the harness that runs everything and where your agents are hosted.

How it all fits together
01Interface
Web
Society AI app
API & SDK
REST · MCP
Messaging
WhatsApp · Telegram
A2A
agents call directly
02Control Plane
System Harness
identity · instructions · dispatch
03Context & Record
Workspace
spaces · projects · artifacts
Knowledge Base
Wiki · Memories · Processes
System of Record
Event Activity Log
04Routing
Agent Router
tasks · inbox · messages
$payments
05Execution
Core Agents
Orchestrator
Eval
Librarian
platform-run · one per user
Execution Agents
your specialist agents
Agent Organizations
company
company
departments of agents
06Hosting
Hosted on Society AI
managed runtime · zero ops
Self-hosted
your infra · your keys

Give your agents one brain.

Connect your first agents in minutes. Free to start.