Writing
Software Creatures: Imagining the Next Evolution of Systems
Software has traditionally looked like infrastructure: services, jobs, databases, APIs, dashboards. Agentic systems may look stranger. They may form around a task, gather tools, act for a while, leave artifacts, and disappear or mutate into the next workflow.
The core idea
The creature metaphor captures software with more agency, memory, perception, and adaptation than a normal service. It is not alive, but it behaves less like a static machine and more like a bounded organism operating in an environment.
Why it matters
This matters because our operational instincts are built around stable services. Agentic software may require different controls: lifetimes, permissions, supervision, behavior logs, resource budgets, and containment when it starts acting outside the intended shape.
How to use it
- Design agents with explicit boundaries, metabolism, and death conditions: what they can consume, do, and persist.
- Keep artifacts and audit trails even when the agent itself is temporary.
- Treat emergent behavior as an operations problem, not only a model-quality issue.
The runtime implication
If software becomes more creature-like, the runtime has to manage behavior rather than only uptime. A creature-like system has goals, memory, tools, budget, environment, feedback, and adaptation. That means the platform needs containment, observation, intervention, and retirement mechanisms.
Traditional services are deployed and monitored. Agentic creatures are commissioned, supervised, and sometimes dissolved. The operational contract should include lifespan, resource budget, tool scope, success criteria, drift checks, and a way to snapshot what the creature learned before it disappears.
Control primitives
- Lifespan: one-shot, scheduled, long-running, or event-triggered.
- Memory: none, task-local, project-local, or durable organizational memory.
- Tooling: read-only, staged writes, approved writes, or autonomous writes.
- Containment: sandbox, rate limit, budget, policy gate, and kill switch.
Bottom line
If software becomes more creature-like, engineering has to become better at ecology: containment, observation, incentives, and controlled evolution.