Skip to content Guth LABS

Agent Engineering

Agent Engineering

Set up one agent with clear records, tests, and an owner who can maintain it.

Planned service · Inquiries open · Scope and availability agreed together.

What this helps with

This Agent Engineering work begins with the detailed setup and indexing needed for a named agent. It makes source ownership, version, sensitivity, and approval visible, then gives the customer owner a chance to build from the approved structure, update, roll back, and run affected tests before handoff.

Good fit

A named, bounded agent has an available owner, existing instructions, and a known failure. A smaller platform setup or ordinary knowledge-base cleanup may be sufficient in some cases.

What you get

Agent profile

Intake, scope, and an approved agent profile.

Source and tool record

A source and tool record with a gap register.

Test results

A direct, paraphrased, stale, conflicting, forbidden, and unanswerable-case test record.

Maintenance handoff

An update and rollback handoff, plus explicit optional-index status.

How it works

  1. Structure: Define the agent, owner, audience, authority, and approval boundaries.
  2. Index: Record instructions, sources, tools, access, and unresolved gaps.
  3. Build and review: Assemble a versioned profile, exercise agreed cases, and hand off update and rollback work.

Details and examples

Full planned process
  1. Identify the agent, version, environment, owner, and use case.
  2. Define its audience, allowed and prohibited behavior, approvals, and unknown handling.
  3. Inventory instructions, sources, tools, and access.
  4. Record source owner, date, sensitivity, and approval.
  5. Flag contradictions, orphaned material, unsupported claims, unavailable sources, and excess permission.
  6. Assemble an approved, versioned profile.
  7. Test direct, paraphrased, stale, conflicting, forbidden, unanswerable, and tool-choice cases.
  8. Let the customer owner perform an update, rollback, and affected tests.
  9. Record the optional index-submission status.
Scope and starting point
Starts with
A named agent, an owner, and its existing materials.
Leaves you with
A versioned profile, source record, build and test record, and maintenance handoff.
Optional
Submission to a developing Guth agent index, without guaranteed admission.
Build and optimize against the same cases.

Compare the current build with proposed changes to its instructions, indexed sources, and tool scope. Keep the test conditions and results with each version, including regressions and unresolved failures, before choosing what to retain.

When to add a full agent audit

Full review service

Need an audit of the agent’s behavior?

Agent setup creates a maintainable foundation. A full agent audit is a separately scoped planned service for reviewing behavior against agreed cases.

Explore full agent audits

Example: PartsHelper-v0.3

Illustrative example

PartsHelper-v0.3

A fictional agent has eight manuals and an ordering tool. The walkthrough identifies inconsistent prices, an ownerless sheet, an unsupported fee, an inaccessible manual, and permission beyond manager approval. These are illustrative issues, not run results.

Questions

Is this public search indexing?

No. Optional submission is to a developing Guth agent index, not Google indexing or guaranteed admission.

Will anyone need passwords?

No passwords need to be handed over. Access and sources are agreed before work begins.

Does this include training, monitoring, or certification?

No. Training, penetration testing, certification, continuous monitoring, and promised answer improvement are outside this planned scope.

Scope, availability, timing, and price are agreed before work starts.

Talk about this service

We can agree the scope before work begins.

Discuss agent setup