agent-platform-skills
Field notes for building on Gemini Enterprise Agent Platform: ADK vs A2A registration, A2UI rendering, Agent Identity, Agent Gateway, what c
Open source Open in the app JSON README (API)
About
Field notes for building on Gemini Enterprise Agent Platform: ADK vs A2A registration, A2UI rendering, Agent Identity, Agent Gateway, what crosses an agent-to-agent hop, and where state lives across Sessions and Memory Bank.
Details
- Kind
- Plugins
- Topic
- AI, RAG & memory
- Publisher
- alanblythe
- Origin
- gemini
- Category
- ferramentas
- Version
- 0.1.0
- Last push
- 2026-08-21T12:23:33Z
- Repository state
- ativo
- License
- NOASSERTION
- Added
- 2026-08-30 14:13:39
- Updated
- 2026-08-30 14:13:39
- Origin id
alanblythe/agent-platform-skills
README
# agent-platform-skills
Skills for building on **Gemini Enterprise Agent Platform** — the parts that are easy
to get wrong and slow to diagnose: which registration mode can render A2UI, what
authenticates on each path, how Agent Identity changes outbound calls, how to tell a
rejected credential from a missing permission, and which store actually keeps data
across restarts and users.
Derived from running deployments. Where a claim was verified by observation rather
than taken from documentation, the skill marks it; where documentation and observed
behaviour disagree, both are recorded.
## Skills
| Skill | Use it for |
| :--- | :--- |
| [`agent-platform-architecture`](skills/agent-platform-architecture/SKILL.md) | **Designing.** The platform constraints that govern how agents fit together, and the six architectures they permit. One agent or several, Agent Identity or not, whether you need Agent Gateway, how to expose an agent in Gemini Enterprise. |
| [`agent-platform-a2a`](skills/agent-platform-a2a/SKILL.md) | **Debugging.** 401 vs 403, identifying the real caller from audit logs, Agent Identity principal forms, reading permissions out of errors, and the conditions under which a broken deployment reports success. |
| [`agent-platform-implementation`](skills/agent-platform-implementation/SKILL.md) | **Building and verifying.** The deployment runs and is quietly wrong: what actually crosses an agent-to-agent hop, why a caller invents data a sub-agent holds, guards a model cannot game, how to build an A2UI surface that survives the round trip, and the verification traps that report a score unrelated to your agent. |
| [`agent-platform-state`](skills/agent-platform-state/SKILL.md) | **Persisting.** Where data lives and what survives: the three stores and their lifetimes, why `app:`-prefixed state does not scope on Agent Runtime, memory scopes and sharing a collection across users, and which store fits which workload. |
The split is design, build, operations, and state, so that none crowds out the
others: the architecture skill states the constraints and the designs they
permit, the implementation skill covers what goes wrong once it runs without
erroring, the debugging skill covers failure modes and how to identify them, and
the state skill covers where data lives and what survives a restart.
## Install
Both installers read the `skills/` directory at the repo root.
```bash
# Claude Code (global)
npx -y skills add https://github.com/alanblythe/agent-platform-skills -y --agent claude-code -g
# Project scope instead of global — drop the -g and run from the project directory
npx -y skills add https://github.com/alanblythe/agent-platform-skills -y --agent claude-code
# Antigravity / Gemini CLI
npx -y skills add https://github.com/alanblythe/agent-platform-skills -y --agent antigravity -g
```
To use it without installing, point your agent at
[`skills/agent-platform-a2a/SKILL.md`](skills/agent-platform-a2a/SKILL.md) directly — it
is a single self-contained Markdown file.
## Layout
```
plugin.json # Claude Code plugin manifest (duplicated in .claude-plugin/)
.claude-plugin/plugin.json
gemini-extension.json # Antigravity / Gemini extension manifest
skills/
README.md
agent-platform-architecture/
agent-platform-implementation/
agent-platform-a2a/
SKILL.md # YAML frontmatter (name, description, metadata) + body
agent-platform-state/
```
Each skill is a directory under `skills/` containing a `SKILL.md` whose frontmatter
carries `name` and a `description` written as trigger phrases — that description is what
an agent matches against when deciding whether the skill is relevant, so it names the
symptoms ("A2UI renders as raw JSON", "401 between agents") rather than summarising the
contents.
## Scope
Deliberately not covered: writing agent code, evaluation methodology, and scaffolding.
Those are handled by the [`google-agents-cli`](https://github.com/google/agents-cli)
skills, which this complements rather than replaces — install both.
The one exception is a section of `agent-platform-implementation`: where the eval
harness reports a number that is not about your agent — a stale trace regraded as
if it were new, a metric grading the transport your users are not served, a
multi-turn case whose earlier turns were narrated rather than run. Choosing metrics,
dataset schema and the flywheel itself remain with `google-agents-cli`.
## License
Apache-2.0