Lifecycle and Orchestration (L)
This page corresponds to §6 of Agent Harness Engineering: A Survey. Lifecycle and orchestration concerns how an agent system carries a task across repeated model calls, tool calls, failures, revisions, handoffs, and completion.
The layer combines two concerns that frameworks often separate: execution flow and the operational state that execution reads and writes.
Lifecycle State
Lifecycle state is the operational state needed to continue a run:
- pending subtasks;
- tool results;
- intermediate artifacts;
- repository changes;
- coordination metadata;
- session persistence;
- checkpoints and resume points.
This differs from the Context layer, which decides what information is shown to the model, and from Observability, which records what happened. Lifecycle state is the state the harness uses to continue and coordinate execution.
The paper distinguishes:
| Model | Description | Tradeoff |
|---|---|---|
| Stateless replay | Reconstruct run state from recorded interaction history. | Reproducible and auditable, but grows expensive with trajectory length. |
| Stateful execution | Store operational state outside the prompt. | Better continuity and recovery, but harder consistency and debugging. |
| Hybrid | Combine replayable logs with external operational state. | Most practical long-running systems land here. |
Three Orchestration Levels
Single-Agent Inner Loop
The basic unit is a single agent repeatedly reasoning, acting, and observing. ReAct is the conceptual primitive, but the behavior is determined by the harness as much as by the model: prompt construction, tool dispatch, control flow, and tool-output feedback all shape the loop.
Examples include Claude Code, Codex CLI, Aider, SWE-agent, Gemini CLI, and OpenCode.
Multi-Agent Orchestration
Multi-agent systems separate planning, execution, checking, and revision across roles. The survey groups patterns such as:
- hierarchical orchestration;
- named teams;
- graph composition;
- workflow orchestration;
- fan-out exploration.
These systems are usually stateful or hybrid because they must track roles, assignments, task graphs, shared artifacts, and coordination state.
Full Lifecycle Pipelines
Issue-to-pull-request systems manage the whole workflow from specification to verified output. The central abstraction is a task runner: scheduling, state persistence, retry, validation, iteration, and delivery over durable artifacts such as repositories, issues, branches, files, tests, and pull requests.
Vibe Kanban, Symphony, and GitHub Agentic Workflows are representative of this level.
Open Problem: Handoff Contracts
The paper argues that modern harnesses distribute work across planners, subagents, tools, sandboxes, evaluators, and humans, but the interfaces between them remain ad hoc.
A reliable handoff should transfer more than a text summary:
- intent;
- constraints;
- permissions;
- artifacts;
- provenance;
- budget state;
- risk level;
- trace history;
- unresolved decisions;
- conditions for returning control.
This is why issue trackers, repositories, and task runners become lifecycle control planes. They preserve state, responsibility, and evidence across agent/human boundaries.