FactoryMind AgentOS

FactoryMind AgentOS is a Python library for building reusable, reliable, self-improving agents with a typed runtime contract.

Current status: v1.3.0 (stable).

Install

pip install agent-os==1.3.0

For local repo usage:

pip install -e .

Stable Core API

• AgentOS
• AgentOS.load(...)
• AgentOS.create_agent(...)
• AgentOS.sessions
• AgentOS.memory
• AgentOS.learning
• AgentOS.flame
• AgentOS.tools
• AgentOS.capabilities
• AgentOS.monitor
• Core models: AgentDefinition, Session, SessionEvent, AgentOutput, MemoryItem, LearningRun, ToolManifest, ToolCallResult, ModelCapability, UsageRecord, CostRecord, AgentStatus, PoolItem, ReflectionBatchRun

Breaking change in v1.2.0: candidate-era learning APIs are removed from stable surface. FLAME temporary-memory reflection is the authoritative learning path.

v1.3.0 adds dual output modes per agent:

• output_mode="text" (default)
• output_mode="json_schema" with output_schema and typed AgentOutput.content_json

Agent Capability Tiers

• basic_agent: short-term session context only, no long-term memory retrieval, no learning
• self_learning_agent: short-term + long-term memory (agent-owned memory only) + learning

Example:

from agent_os import AgentTier

app.create_agent(
    agent_id="project-selector",
    goal="Select best project",
    model="gpt-4.1-mini",
    agent_tier=AgentTier.BASIC_AGENT,
)

Library Quickstart

from pathlib import Path
from agent_os import AgentOS

root = Path(".agent-os")
app = AgentOS.load(root=root, runtime_mode="local")
app.create_agent(
    agent_id="proposal-agent",
    goal="Draft project proposals",
    model="gpt-4.1-mini",
    tenant_id="default",
    output_mode="text",
)

session = app.sessions.init("proposal-agent", "Draft a proposal for Project Orion")
output = app.sessions.run(session.session_id)
print(output.type.value, output.content)

app.sessions.feedback(session.session_id, "Make it concise and include delivery timeline.")
app.sessions.accept(session.session_id, note="Accepted")

Learning and Tools

# learning.run remains as a compatibility alias and dispatches to FLAME trigger.
run = app.learning.run(agent_id="proposal-agent", window_size=20)
runs = app.flame.list_runs("proposal-agent")
pool = app.flame.list_pool("proposal-agent")

from agent_os import ToolManifest, ToolScope

tool = app.tools.register(
    ToolManifest(
        name="project_lookup",
        scope=ToolScope.READ,
        input_schema={"type": "object", "properties": {"q": {"type": "string"}}, "required": ["q"]},
        output_schema={"type": "object"},
    )
)
app.tools.bind("proposal-agent", tool.tool_id)
result, audit = app.tools.call("proposal-agent", session.session_id, tool.tool_id, {"q": "orion"})

Canonical Examples

• examples/proposal_agent_app.py
• examples/project_selection_agent_app.py
• examples/keyword_extraction_agent_app.py

v1 Docs

• docs/stable-sdk-contract.md
• docs/migration-v1-memory-only.md
• docs/migration-v1.2.0-flame-cutover.md
• docs/migration-v1.3.0-structured-output.md