Architecture⚓︎
ContextAgent centers on a context core that every agent connects to. Pipelines orchestrate the core, agents execute tasks through LLM providers, and tools expose capabilities back to the core.
System layers⚓︎
- Pipeline runtime — entry point that maps queries to structured runs, handles retries, and stitches together steps.
- Context core — stores prompt templates, run state, artifacts, and memory so that every agent acts on shared truth.
- Agents — lightweight wrappers over LLM clients; each agent references a template in the context core.
- Tools — external capabilities (search, code execution, vector DB) registered with the context core and gated via policies.
- Observability — auto-traced spans, structured logs, and artifacts pushed to disks or remote stores.
graph LR
subgraph Client
Q[Pipeline Query]
end
Q -->|Validates| R[Pipeline Runtime]
R -->|Mutates| C[(Context Core)]
C -->|Selects| A1[Task Agent]
C -->|Selects| A2[Reviewer Agent]
C -->|Dispatches| T[Tool Bus]
T --> S[(Services)]
A1 --> M1[(LLM Provider)]
A2 --> M2[(LLM Provider)]
C --> O[Observability]
Configuration model⚓︎
- YAML pipeline configs define agents, prompt templates, default models, tool access, and tracing settings.
- Profiles expose reusable bundles of context so you can version prompts independently of pipelines.
- Artifacts capture intermediate datasets, summaries, or generated files; the context core makes them discoverable per run.
Recommended folder layout
pipelines/configs/*.yaml— pipeline definitionscontextagent/profiles/— prompt templates + state defaultscontextagent/tools/— structured tool definitionsdata/— input datasets or assets consumed by pipelines
Deployment topology⚓︎
| Mode | Description | Best for |
|---|---|---|
| Local | Single process running pipelines on demand | Prototyping, testing |
| Worker Fleet | Pipelines dispatched to async workers via queues | Production workloads |
| Hybrid | Local for experiments, remote for heavy agents | Teams with mixed needs |
Continue with the Agent Lifecycle to understand how agents progress during a run.