Beyond Multi-Agent Collaboration: State Ownership in Automated Pentesting
Why execution boundaries, durable state, and work-unit design matter more than multiplying agents.
Multi-agent collaboration is often treated as the default answer to long-running automation. In penetration testing, that assumption deserves more scrutiny. A campaign is not simply a collection of independent questions: it is a sequence of observations, hypotheses, validations, and actions whose meaning depends on prior state.
The campaign is a stateful investigation
A useful execution model must preserve what has been observed, what has been tested, which evidence supports a finding, and which actions have already changed the target environment. Parallelism is valuable only when those boundaries are explicit.
Parallelize independent work, not shared uncertainty
Discovery against separate assets, passive enrichment, and independent evidence review can often run concurrently. Exploitation chains that depend on session state, access changes, or a previous result are much harder to parallelize safely.
The right unit of work is therefore not necessarily an agent or a service. It is a bounded operation with explicit inputs, outputs, preconditions, side effects, and a recovery strategy.
Throughput needs an execution model
A single worker may be too slow, but adding workers without ownership rules can create duplicate actions, stale assumptions, and broken evidence chains. A scheduler should understand dependencies and resource constraints, while a journal records durable progress.
The objective is not maximum concurrency. It is higher throughput without losing the causal history that makes a security result defensible.