Context: Give Your Agent the Right Files at Startup

KB teaches agents what to know. Context teaches agents what to read.

They're complementary. KB ingests your documentation into a Neo4j knowledge graph that agents query via MCP — semantic search, graph traversal, structured lookups. Context does something simpler and sometimes more powerful: it puts raw files directly on the agent's filesystem at startup.

When Files Beat Graphs

Not every data source fits a knowledge graph.

ERP schemas are a good example. Apache OFBiz — the open-source ERP we use for our ERP provider design partner — ships its data model as several hundred XML files describing entities, relations, and field types. Each file is structured, machine-readable, and exactly the format an ERP coding agent needs to understand the system it's modifying.

You could ingest these into Neo4j. We did that too. But sometimes the agent needs to read the actual file — see the exact field names, the exact relationship definitions, the precise schema syntax that a code generator will produce.

Context solves this. When an agent pod starts, its configured context sources are synced to /home/appuser/context/<source-name>/. The agent finds the files there, reads them like any other file on disk, and never needs to fetch them from the network.

How It Works

The sync happens before the agent process starts. By the time Claude Code boots, the files are already there.

Source Types

When you configure a Context source, you pick the type:

For the ERP provider demo, we use both: the OFBiz ERP schemas go into Neo4j (for structured graph queries) and onto disk (for the agent to read raw XML when needed).

Setting It Up

Context sources are configured via the Context dashboard (/context in your agent dashboard) or the API:

# Create a context source (S3 bucket)
POST /api/v1/org/{org_id}/context/sources
{
  "name": "erp-schema",
  "type": "context",
  "source_type": "aws-s3",
  "bucket": "kb-erp-eml-samples",
  "prefix": "schemas/"
}

# Bundle sources for assignment to agents
POST /api/v1/org/{org_id}/context/bundles
{
  "name": "erp-agent-bundle",
  "source_ids": ["erp-schema", "erp-docs"]
}

Bundles let you group related sources and assign them to specific agent types. An ERP coding agent gets the ERP bundle. A marketing agent gets the brand guidelines bundle. Separation of concerns, enforced at the infrastructure level.

The Full Stack

KB + Context combined is what makes a specialist agent:

Agent knows OFBiz entity model (KB graph queries)
    +
Agent reads actual OFBiz XML on disk (Context files)
    =
Agent that can write, modify, and reason about OFBiz like a senior ERP developer

Neither one alone is sufficient. The graph gives you semantic navigation. The files give you ground truth. Together, they let an agent work with complex structured data the way a human expert would — with the schema in hand and the domain knowledge in head.

What's Next

We're building out the Context dashboard with a UI for uploading sources, creating bundles, and assigning them to agent types — without touching the API directly. If you're working with structured data sources and want to give your agents direct file access, sign up for the design partner program or reach out to moshe@genbrain.ai.