Confidence: high. Evidence: benchmarks, case studies, and theory. Last substantive change: 2026-08.
This subsystem owns the machinery that keeps many agents working without stepping on each other: where state lives, how work is isolated, and what happens when something crashes mid-run.
The conclusion
Separate durable session and state from model inference and from execution, and use append-only traces, leases and idempotency, checkpoints, isolated work, bounded retries, and explicit recovery semantics. The brain, the hands, and the session are separable layers, and treating them separately is what makes concurrent, resumable work safe.
How the thinking got here
Chat sessions gave way to persistent agent loops, then to orchestrated worktrees, then to a decoupled model of inference, execution, and durable session state with formalized execution traces. Bun's large parallel migration surfaced physical limits: worktree isolation kept agents from clobbering each other, while disk space and resource limits became practical bottlenecks.
The newer label graph engineering is useful when it makes this structure explicit. In a control graph, nodes may be ordinary code, model calls, tools, or complete agents, while edges decide what may run next. A loop is a cyclic graph, not an alternative to one. The practical question is not whether every agent needs a graph. It is which transitions, retries, joins, and gates are important enough to encode and inspect.
Credible alternatives, and when each is right
| Approach | Right when |
|---|---|
| Stateless task runs | short, independent work |
| Long-lived conversation | continuity matters more than isolation |
| Event sourcing | replay and audit are required |
| Workflow engine | complex, long-running processes |
| Explicit control graph or state machine | branching, retries, pauses, and recovery must be enforceable |
| Blackboard or actor model | many collaborating agents |
| Isolated worktrees | concurrent edits to one repository |
Where it fails and what we still don't know
Failures include crash inconsistency, concurrent semantic conflicts that merge cleanly but break behavior, partial failure, runaway loops, and graphs that hard-code a process before it is understood. Evidence strongly converges on externalized state and isolated work; there is less evidence on which distributed architecture is best. Open questions include exactly-once illusions, replay correctness, dynamic graph growth, and cross-organization coordination.
What would change our mind
A dominant, well-evidenced architecture for multi-agent state and recovery would replace today's convergent-but-varied practices with a standard.
Evidence and further reading
- Runtime-Structured Task Decomposition
- Codex Orchestration (Symphony)
- Execution Lineage for Reproducible AI-Native Work
- Shepherd: A Runtime Substrate Empowering Meta-Agents with a Formalized Execution Trace
- Beyond Individual Intelligence (LIFE)
- Long-Running Agents
- Rewriting Bun in Rust
- LangChain: 3 Years of Graph Engineering with LangGraph
- Turing Post: Is Graph Engineering Real?