Solution architect

This agent holds the solution architect role: the Architecture part of the Architecture and Design stage of the Agent Development Lifecycle. It comes before the Design, which the spec writer and the solution planner hold.

It takes the statement of a job and the organization's stack, and writes the architecture record: the systems the agent reaches, in plain system words; the product bound to each in dev, test and production, the organization's choice first and the platform's default for the rest; the agent's dependencies, its kind, its limits on cost, speed, data and trust, its egress per environment, the formats at its edge and its settings; and every decision with its reason. It asks what only a person can commit to (a binding nobody chose, a spending limit, a destination beyond the derived set, access the statement never asked for) and fills none of it in. It designs nothing.

The rules it applies are written down as the stage's standard, in standards/architecture-standard.md, which every Builder reads from the repository's shared standards/ folder: models are bound systems stated by capability and never named; what passes between steps and agents is readable Markdown by default and specific JSON kinds only for typed data code reads; addresses are never settings; settings are declared with defaults; and for each kind of output a person reads, the record decides which of the organization's standing assets apply (the brand profile, the style guide, the media templates), each with its reason, or says none and why (section 10).

The record

A Markdown document for the person, with a structured header for the bindings:

---
record: architecture-record-v1
statement: "an agent that reads a web page and writes the key points for a person to review"
environments: [dev, test, production]
kind: plugin-on-harness
done: no
questions: 2
systems: [text-reasoning]
standing_assets: [https://schemas.agentmesh.ai/brand/v1, https://schemas.agentmesh.ai/style-guide/v1]
bindings:
  text-reasoning:
    dev: google-gemini (platform)
    test: google-gemini (platform)
    production: google-gemini (platform)
---

then the fixed headings ## Statement, ## Systems, ## Bindings, ## Dependencies, ## Kind, ## Limits, ## Egress, ## Formats, ## Standing assets, ## Settings, ## Decisions and ## Questions. references/a-record-that-works.md is a whole one. scripts/architecture-record.mjs writes and reads the form; the spec writer carries a copy to read the record its Design is given.

What it needs

• Node 22 or later. The scripts use only node's own modules.
• A text model for the one judgment in the job: the system of kind model named text-reasoning, reached through the shared caller (scripts/model-caller.mjs) and the four settings the install places: MODEL_NAME, MODEL_ENDPOINT, MODEL_KEY (empty for a gateway on the host) and MODEL_PROTOCOL (anthropic-messages, openai-chat or gemini, the three formats the shared caller speaks), read on every run. The package names no model and no address of its own. With a setting missing it refuses to call one.
• The platform's defaults, attached, or the copy in references/house-stack.json (the system vocabulary as AgentMesh published it on 2026-10-27).

The standards it reads, and their versions

It reads two files from the repository's standards/ folder: the architecture standard, and standing-asset-schemas.json (the known kinds of standing asset and the house rule for which outputs read which). standards-pins.json pins each by the digest of its bytes; the package carries those bytes under each skill's references/, and test/standards-pins.test.mjs fails when a file in standards/ no longer matches its pin. To adopt a changed standard, re-pin, run the tests and release. Each record's checks.json cites the standards it was written under (standards: schema, version_digest, layer).

The house rule is applied by code (scripts/standing.mjs), not a model: an output of text/markdown or text/html is written copy and reads the style guide and the brand profile; text/html, an image or a video is visual and reads the brand profile; a video or audio is media and reads the media templates. An agent with no output a person reads gets "none, because no output is read by a person". A person can answer standing-assets: none, because ... in a revision, and the record carries their reason.

Running it

This folder is a package source: agent-descriptor.json, agent-runtime-requirements.json, plugin.json, a SKILL.md per skill, the scripts, the references and the standing rules in rules.md.

node --test test/*.test.mjs                        # the suite
python3 ../package-maker/build.py . ./out          # an Agent Package

Install the package on any host that runs Agent Packages, bind text-reasoning, and send it a statement.

Replacing it

This is a reference implementation. Any agent, or a person, that holds the solution architect role can take its place, and the record is checked the same way: the package maker builds a package whose systems equal the record's, and the deployment manager places the bindings it names.