Agent

An agent is Tyrum's durable runtime persona: one identity boundary that keeps conversations, memory, work state, and policy coherent across surfaces and turns.

Read this page

• Read this if: you want the top-level mental model for how one Tyrum agent stays coherent over time.
• Skip this if: you already know the agent boundary and need mechanics details.
• Go deeper: use Messages and Conversations, Conversations and Turns, Memory, Work board and delegated execution, and ARCH-20 conversation and turn clean-break decision.

Agent subsystem map

Agent boundary

What the agent owns

• Agent-scoped runtime configuration: identity, tone, tools, MCP, skills, memory policy, and prompt shaping.
• Durable continuity through conversations and conversation-scoped context assembly.
• Turn orchestration: prompt assembly, model calls, tool use, and durable follow-up decisions.
• Memory retrieval and writes that preserve long-term continuity across surfaces.
• WorkBoard updates that keep background commitments explicit instead of relying on transcript recall.

What the agent does not own

• Transport, protocol validation, and edge connectivity ownership.
• Human approval policy or authz policy authoring.
• Node-local execution internals or provider-specific channel behavior.

Primary flows

Interactive turn flow

1. A surface event is routed to the agent through a conversation.
2. The runtime assembles prompt context from conversation state, memory, and active work state.
3. The turn produces progress, durable state updates, and any needed follow-up turns or child conversations.

Background progress flow

1. Work or automation targets the same agent through a dedicated conversation.
2. The agent continues making progress through durable turns under the same conversation and state model.
3. Outcomes are reflected back into transcript, conversation state, memory, and WorkBoard.

Invariants for this boundary

• Agent continuity is scoped by agent_id.
• Memory is agent-scoped and survives conversation boundaries.
• Conversation context is partitioned by conversation_id, not by implicit hidden sub-contexts.
• Turns are serialized per conversation so one context boundary has one active line of reasoning at a time.
• Durable state, not transcript replay alone, is the recovery source of truth.

Go deeper

• Architecture
• Messages and Conversations
• Conversations and Turns
• Transcript, Conversation State, and Prompt Context
• Channels
• Memory
• Context, Compaction, and Pruning
• Work board and delegated execution
• ARCH-20 conversation and turn clean-break decision
• System Prompt
• Multi-Agent Routing
• Agent Loop