Agents & multi-agent

What an Agent is

An Agent is the basic unit that executes work on the platform. Each Agent has:

Role & system prompt — defines goals and behavioral constraints
Callable tools — built-in primitives + Skills currently activated
Context — session history, memory retrieval, external materials
Model backend — routes tasks to appropriate LLM providers
Ledger — step-by-step record of decisions, tool calls, results, and confidence

Two Agent types

Two kinds of Agents run on the platform, each with a distinct role:

Type	Role	Mechanism
Orchestrator	Understands tasks, splits work, schedules child Agents, decides completion	Dual loop (outer: Epoch / Plan + inner: Round / Ledger)
Assistant (sub-Agent)	Executes assigned subtasks	Parallel spawn supported; isolated context each; results returned to Orchestrator

Orchestrator dual loop

Rendering diagram…

Outer loop — formulates / adjusts the Plan and stage goals
Inner loop — advances multiple Rounds inside one Plan stage; each Round writes to the Ledger
Ledger — every step is recorded; supports replay / audit / verification afterward

Parallel sub-agents

For complex tasks the Orchestrator spawns multiple Assistants in parallel:

Each Assistant has its own context (saves tokens on the main Agent)
They are mutually invisible (no cross-talk)
Results are returned to the Orchestrator in a structured form

Typical cases:

Parse 10 documents at once → 10 parallel Assistants
Large-project traceability: pull requirement modules in parallel → build candidate links each → Orchestrator merges

Multi-agent coordination

Mechanism	Description
Delegation	Top-down task assignment; Orchestrator → Assistant
Grafting	Reference outputs across parallel sessions (see Graft skill)
Hook coordination	Shared constraints injected at key steps via Hooks
Memory sharing	Cross-Agent KB / LTM via MemoryOS

Ledger: audit and replay

The Ledger is the Orchestrator’s execution log; each step records one row:

[Round 3] decide → tool_call(alm.import, {project: "ABZ"})
           confidence: 0.87
           trace_id: abc-123
           result: 1859 items imported
           cost: 12.3s / 4200 tokens
[Round 4] decide → delegate(trace-sub-agent)
           confidence: 0.92
           ...

Uses:

Post-hoc audit — reconstruct each step when something breaks
Cost accounting — token / duration per Round
Self-check — low-confidence steps trigger backtracking or user confirmation

Session — Ledger is attached to a Session
Skill system — extensible Agent capabilities
Graft
Observability — how Ledger is visualized
Code entrypoints: AgentFlow/src/orchestrator/, AgentFlow/src/assistant_agent/