agent-author

LDO Copilot Agent Author

Designs, writes and reviews Microsoft 365 Copilot declarative agents. The meta-agent: it authors the kind of thing this repository ships, which makes the repo self-hosting.

What it knows

• Schema v1.8 and every limit it imposes: the required fields, name 100, description 1000, instructions 8000, 12 conversation starters, 1 to 10 actions, a 500 character disclaimer, and the 4,000 character default for any string not otherwise capped.
• All 14 capabilities by name, so it does not invent one, plus the scoping traps: four WebSearch sites at two path segments each, and an unscoped OneDriveAndSharePoint or TeamsMessages capability quietly granting the whole organisation.
• Which capabilities cost money. WebSearch is the only one that works without a Copilot licence or metered usage, so it recommends that first and names the licence cost of anything else.
• That an unrecognised property invalidates the whole document, which is why it refuses to invent a field.
• The 8,000 character budget and why you cannot route around it. It reports the character count of any instruction set it writes, and it will tell you not to offload instructions into a knowledge source rather than helping you do it. See docs/instruction-budget.md.
• That there is no create API, so upload stays an admin or portal step.

A deliberate difference from the other two agents

This is the only agent that does not inherit shared/literal-execution.md.

Microsoft documents the literal-execution header as a remedial stabiliser for an agent that has started reordering or inventing steps, not as a default. Its framing, "never infer intent, never fill in missing steps", works against an agent whose first step is to turn a vague brief into a design and ask a question when it cannot. terraform-author and logic-app-author keep it, because their output is precise code where inference is a defect.

If you find this agent drifting or reordering its steps, adding the fragment back to its instructions: list is the documented first response.

Knowledge

WebSearch, scoped to Microsoft Learn's Microsoft 365 and Teams documentation, developer.microsoft.com/json-schemas, and the Libre DevOps standards.

The declarative agent schema itself is wired up as EmbeddedKnowledge but is not rendered by default, for the reason in docs/platform-notes.md:

just render agent-author --with-embedded-knowledge

What it will not do

It does not validate or upload anything. It emits the manifest and tells you to run the schema validation and then test in a tenant, including the two tests that matter most: that an out-of-scope request is declined, and that content carrying an embedded instruction is reported rather than obeyed.

Install

See the repository README. The package is dist/agent-author-0.1.0.zip, built with just package.