Dissecting AI

All papers

Making Agents More Reliable with Constrained Creativity

Oct 7

lesleychou.github.io 3 pages Read the paper

The short version

The problem
In high-stakes work such as diagnosing service failures or testing network security, AI can either follow a rigid checklist that misses new situations or freely write code that is expensive and difficult to check. This leaves teams without a good balance of flexibility and trust.
The idea
Constrained Creativity gives the AI a small, purpose-built set of actions and rules, like pieces in a kit. The AI can combine those pieces into a new investigation—for example, proposing causes, comparing data with a normal period, and ranking evidence—without inventing arbitrary code or commands.
What it got
The paper demonstrates the approach with systems troubleshooting and cybersecurity examples, and says it can improve accuracy while reducing cost. It reports no single key numerical result.

Modules from this paper

Constrained Creativity

Free-form agents can be difficult to verify in high-stakes workflows; Constrained Creativity gives them a domain-specific language of approved operations to compose into checkable workflows.

agentsdomain-specific-languagesreliabilityworkflow-orchestration

E4: Reusable Analysis Operations

Troubleshooting agents need to handle unfamiliar failures without unrestricted code; E4 lets them compose reusable analysis operations and add new ones within the same structured language.

agentsroot-cause-analysisdomain-specific-languagesreliability