Releases: eightx-labs/eam-spec
EAM v0.1.0 — Initial Publication
First public release of the EightX Agent Manifest specification.
EAM is a declarative YAML format for defining AI agents independently of the model provider or runtime that executes them. The spec is Apache 2.0; reference validators and compilers are MIT and live in separate repositories.
What's in v0.1.0
- Normative spec (
v0.1/SPEC.md) — RFC 2119 grammar covering identity, skills, connectors, sub-agents, runtime compatibility, and policies. - JSON Schema (
v0.1/schema.json) — Draft 2020-12, strict mode. Validates every example shipped here. - The lock-in surface —
agent.dependencies[]andagent.fine_tuned_ondeclare provider-specific features and fine-tuning that don't survive a runtime migration. Reserved feature identifiers:computer_use,code_interpreter,video_understanding,image_generation,long_context_2m,assistants_threads,file_search_managed. - Four-band runtime model —
preferred/supported/degraded/unsupported. Compilers MUST refuse compilation when a target falls outside the supported set. - Compilation contract (§13) — defines what a conforming compiler MUST emit, including the
compile_report.jsonartefact. - Three reference manifests —
minimal.eam.yaml,pitch_book.eam.yaml, andcassie.eam.yaml(a live production agent illustrating a hard-severity dependency declaration). - Three supporting docs —
docs/portability.md(what survives a runtime change),docs/design-principles.md,docs/compile-targets.md.
Design intent
EAM consumes existing open standards rather than competing with them. Connectors reference MCP servers; inter-agent messaging uses A2A peer references; observability emits via OpenTelemetry.
The spec is deliberately conservative about portability — see docs/portability.md for an honest enumeration of what survives a runtime change and what does not.
Status
v0.x is a stability disclaimer: breaking changes are permitted between minor versions. v1.0 will commit to stability semantics. See §14 of the spec.
The compile-to-Claude reference compiler is in beta. Compile-to-OpenAI and compile-to-Gemini are on the published roadmap.
Feedback welcome
The maintainers are particularly interested in:
- Real-world cases the four-band runtime model can't express.
- Provider-specific features that warrant addition to the reserved feature list.
- Policy declarations that real deployments need beyond the v0.1 set.
Open an issue or pull request. See CONTRIBUTING.md.