Edit: vault-backup/sops/agent-build-loop.md

Note Content (Markdown)

entities:
- - 'SOP: Agent Build Loop'
  - Page
topics:
- agent-build-loop
- sops
- knowledge

# SOP: Agent Build Loop

## Purpose

Create future Hermes agents with enough context, gates and evaluation that they
behave like useful teammates rather than loose prompts.

## When To Use

Use this SOP whenever Anthony or the COO proposes a new agent, workflow or
recurring automation.

## Process

1. Name the job in one sentence.
2. Identify the operating problem.
3. List source systems and evidence.
4. Define what the agent must read, write, draft, refuse and escalate.
5. Define memory needs using the memory filter.
6. Define hard gates.
7. Create role files:
   - `soul.md`
   - `heartbeat.md`
   - `gates.md`
   - role SOPs
   - evals
8. Create 10-20 dry-run eval questions.
9. Test on synthetic or approved notes.
10. Review errors and update the role before any live connector.
11. Log what was learned.

## Output Template

```markdown
# Agent Build Brief: <agent-name>

## Job

## Operating Problem

## Source Systems

## Memory Needs

## Allowed Outputs

## Non-Responsibilities

## Hard Gates

## Evals

## Dry-Run Result

## Approval Needed Before Live Use
```

## Stop Conditions

Stop and ask Anthony if:

- the agent would send externally;
- the agent would change spend, access, credentials, legal, finance or client
  records;
- the source system is unclear;
- the proposed memory requires raw artifact hoarding;
- the role is too broad to evaluate.