POHEEngineering field guide
Field guide № 01Proposed pattern discipline

Pattern-Oriented
Harness Engineering:Design needs Patterns, when Agentic AI meets OOD

A field guide to the recurring decisions around AI agents: what may vary, what must hold, and what happens when it does not.

Browse the 52 records Read the method
A-squared taxonomy: three activities—design, verification, and evolution—cross protected contracts and permitted variation, with an authorized re-generalization return from maintenance to design.
A² / THE TAXONOMY

Two axes, one record.

Activities ask when an obligation is designed, verified, and revisited. Assets ask what must hold and what may change. The dashed return is re-generalization: the contract itself is reopened.

Open the full map Download PDF
52candidate records
05source families
03activity coordinates
Inventory size, not a prevalence claim.
05 / SOURCE FAMILIES
PCoding-agent mechanismsExecution loops, tools, authority, durable state, and stop-time checks.29 recordsWWorkflows & coordinationNamed workflows and coordination forms from official agent guides.7 recordsCCommunicationClassical communication structures applied to agent systems.4 recordsRReasoning & feedbackReason–act and verbal-feedback mechanisms.2 recordsEHarness evolutionResearch mechanisms that adapt or search over harnesses.10 records

Families record provenance, not another taxonomy axis. E records are research mechanisms, not claims of independently established recurring patterns.

H1—H3 / RESEARCH HYPOTHESES

Three hypotheses, one catalog.

Each record informs a hypothesis through its A² activity coordinate: A1 (design) specifies the durable commitment, A2 (verification) exercises mechanized separation, and A3 (evolution) tests bounded self-modification. This is a reading lens—not a change to any frozen pointcut, contract, cue, or evidence value.

H1
Asset hypothesis

Stable assets are design commitments.

Under rapid prompt/code rewriting, the durable asset is an inspectable design commitment: pointcut, contract, permitted variation, test obligation, negative control.

A1 · Design & implementation31 records
H2
Agency hypothesis

Separation must be mechanized.

Agent-led evolution requires guards, gates, checkpoints, rollback, and evaluation that separate proposal from authorization—documentation alone is not a constraint.

A2 · Testing & verification16 records
H3
Controlled-variation hypothesis

Bound the change to clarify the outcome.

Bounded self-modification should reduce contract drift and regression and improve recovery and attribution relative to unconstrained revision.

A3 · Maintenance & evolution22 records
Read the hypotheses and their falsification boundaries
INDEX / 001—052

Pattern catalog

Search by responsibility, mechanism, or join point. Filter without losing the underlying record.

Showing 52 of 52 records

Source anchors are exemplars, not population estimates; coverage is not confirmed prevalence.