Agent Runtime

Chronologue treats time not just as a resource to manage, but as an API for managing and coordinating actions.

The agent orchestration interface is designed to help users audit, steer, and schedule agent behavior directly through their calendar.

This system makes agent actions controllable, and programmable, giving users an interface to assign responsibility to agents in the same way one might call a ride, schedule a delivery, or block time with a collaborator.

Unlike passive logging or traditional chat interfaces, Chronologue anchors agent actions in time with control over when and how long a given action should take.

This introduces precision into task specification and empowers users to compose systems of intent with temporal constraints.

Chronologue reimagines the calendar as a control surface for managing agentic workflows. As language models gain more general capabilities, there is an expanding and implicit need for interfaces that make model behaviors legible, transparent, and editable. Chronologue is built to meet that need.

---

1. System Roles and Components

• User: Initiates prompts, reviews or edits agent-generated plans, and gives feedback.
• Agent: Generates structured plans (e.g., add_event, reflect, reschedule_event) using model outputs.
• Planner Queue: Holds proposed agent actions awaiting user confirmation.
• Executor: Commits approved plans into persistent memory and/or calendar.
• Trace Log: Immutable record of both agent proposals and user decisions.

In the future, everyone will build and manage agents (with varying degrees of precision and scale). Access to tools and information will be ubiquitous — but not all interfaces will confer the same leverage to users.

Agent orchestration is how we build utility and autonomy for humans in a world where the vast majority of materially-impactful work is performed by agents.

---

2. Agent Action Lifecycle

Every agent action flows through a defined lifecycle to support transparency, modification, and feedback:

• proposal: User or agent generates an initial plan, often via function calling from an LLM.
• review: Plan is displayed to the user or organization for inspection or editing.
• approval: User confirms the plan or modifies it to fit their intent.
• execution: Approved plan is committed to memory and/or calendar.
• feedback: Optional user input (rating, revision, comment) is logged to guide future behavior.

Status Transitions

• proposed → approved → executed
• proposed → rejected
• executed → edited

Each transition is recorded as a memory trace, preserving full behavioral provenance.

---

3. Trace Structure for Orchestration

Agent behavior is recorded and governed via structured memory traces. A special type of trace, agent_plan, extends the general schema with orchestration-specific fields.

Required Fields

Field	Type	Description
action_type	string	One of: schedule, reflect, reschedule, log
status	string	One of: proposed, approved, rejected, executed
origin	string	One of: agent, user
scheduled_for	string	Optional ISO timestamp representing intended execution time
uid	string	Unique identifier for trace reference

Trace Linking

• agent_plan traces may link to calendar_event or reflection traces that result from their execution.
• Use fields like linked_event_uid or child_trace_ids to maintain referential integrity.

---

4. Orchestration UI Patterns

Chronologue includes several interface modules to manage agent behavior interactively:

• Daily Agent Preview: Shows all pending proposed actions, with accept/reject controls.
• Command Log: Timeline of all agent outputs and user decisions, chronologically sorted.
• Editable Plan Workspace: UI that allows drag-and-drop reordering, rescheduling, or editing of proposed actions.
• Feedback Modal: Allows users to rate or comment on specific agent behaviors, improving downstream suggestions.

These patterns ensure agents are observable and influenceable—core requirements for high-trust collaboration.

---

5. Feedback Loop, Behavior Adaptation, RL, Federated Learning

The orchestration system supports structured feedback and adaptation.

• feedback_rating and notes can be applied to any trace.
• Feedback can be linked to the agent_plan UID to trace behavioral outcomes.
• Longitudinal analysis of approved/rejected traces can drive:
  • Reward modeling
  • Reinforcement learning (e.g., via success/failure tags)
  • Federated updates to personalized planning models

Feedback is not ephemeral—it’s written to memory for reuse.

---

6. Scheduling and User Customization

User constraints are enforced and communicated to agents through scheduling metadata:

• Time as a constraint interface: Only allow agents to propose actions during unblocked slots.
• User-defined cadence: Set rules for how often agents can propose certain actions (e.g., 1 reflection per day).
• Discrete allocation: Agents may be limited to a budget of actions per day or week, modifiable by the user.
• Prompt context injection: Availability and user preferences are embedded in model inputs.

---

7. Implementation Notes

Chronologue uses the following system architecture for orchestration:

FastAPI Endpoints

• POST /agent/plan — submit an agent proposal
• GET /agent/queue — retrieve pending plans
• POST /agent/approve — approve a specific plan
• POST /agent/reject — reject a specific plan
• POST /agent/feedback — log feedback to a specific trace

Frontend Sync

• Use WebSockets or polling to synchronize planner queue and user approvals in real time.

System Separation

Modular separation is recommended:

• Memory Trace Store: All long-term memory and behavior logs
• Calendar Event Store: Ground-truth schedule and event registry
• Agent Queue Store: Temporary buffer for proposals pending execution

This separation enables caching, rollback, and traceability of agent behavior at different lifecycle stages.