What it is

The main engine that powers the Work agent. The idea behind the harness: make sure that whatever the agent says is correct — validated against a list of deterministic validation checks, written in SQL or Cypher. The current implementation runs on Postgres + AGE.

Because the Work agent does work for a particular, predetermined industry domain, those deterministic evaluations can be generated automatically from the world-knowledge context of that domain. They then run during agent execution, to ensure that whatever decision the agent proposes is correct.

For example: asked to generate a response to an RFP, the Work agent might propose that the work can be done in 60 days with 10 people. Without the harness, nobody catches it when that expectation is outrageously incorrect and deviates from reality. The harness catches it.

In the manifesto's terms, this is where the modal machine disturbance — the confident, wrong proposal — is caught inside the system instead of on a professional's desk.

Architecture

Architecture diagram: the neuro-symbolic AI loop harness for decision intelligence, with assembler, parser, compiler, evaluator and session around an interchangeable CLI agent
The loop. Deterministic validation of AI output with ontological, semantic and analytical modelling. The agent is interchangeable; the harness and its validations are not.