Context engine

The context engine is the authoritative source of truth for prompts, state, and artifacts. It ensures that every agent sees the same world, no matter which pipeline invokes it.

Primitives

• Profiles — typed templates containing reusable system + assistant prompts.
• State — mutable key-value store scoped by run, agent, or tool.
• Artifacts — rich payloads (files, embeddings, structured JSON) referenced by agents downstream.
• Events — timestamped records emitted during a run for tracing and analytics.

from contextagent.context import Context

context = Context(["profiles", "state"])

context.set_state("analysis.notes", "Focus on customer sentiment and product gaps.")
profile = context.get_profile("data_science/analyst")
context.log_event(kind="checkpoint", data={"step": "analysis"})

Access patterns

Operation	Method	Notes
Read profile	context.get_profile(name)	Returns structured config objects
Update state	context.set_state(path, value)	Path-based addressing with dot notation
Fetch artifacts	context.get_artifact(name)	Supports lazy loading for large files

Use state scopes to keep data isolated:

• run.* for orchestration-wide metadata.
• agent.* for specific agent instructions or scratchpads.
• tool.* for tool outputs and caches.

Versioning

Profiles and templates can be versioned individually. Example directory layout:

contextagent/
  profiles/
    data_science/
      analyst.yaml
      analyst@v2.yaml
    support/
      triage.yaml

Set the active revision in pipeline configs:

agents:
  analyst:
    profile: data_science/analyst@v2

Persistence

• In-memory for local development.
• Redis / Postgres adapters for distributed runs.
• Artifact stores (S3, GCS, local disk) configured per environment.

Custom adapters

Implement ContextStore and ArtifactStore interfaces to plug in your own persistence layer.

With the context engine in place, explore tooling integrations to extend agent capabilities.