Context Engineering for Agentic Systems

Prompt engineering got all the attention first.

Context engineering is what matters when the agent touches real work.

An agent does not need every document, every Slack thread, every ticket, and every log line. It needs the right context, at the right time, with the right permissions, in the right shape.

That is an architecture problem.

Context is not just retrieval#

Context can come from:

• User input.
• Workflow state.
• Tool results.
• Docs.
• Tickets.
• Code.
• Logs.
• Traces.
• Metrics.
• Memory.
• Policies.
• Prior approvals.

Treat these as separate sources with different trust levels.

Freshness matters#

Old docs can be worse than no docs.

Every context source needs a freshness story:

• When was it updated?
• Is it authoritative?
• Does live system state override it?
• Should stale data be labeled?
• Should the agent ask for confirmation?

A runbook from six months ago should not outrank a failing trace from five minutes ago.

Scope context before retrieval#

Do not retrieve first and filter later.

Scope by:

• User permissions.
• Workspace.
• Customer.
• Service.
• Repository.
• Incident.
• Data classification.
• Time window.

This keeps context engineering tied to security.

Compress without erasing evidence#

Long context is tempting. It is also expensive and noisy.

Use compression carefully:

• Keep source links.
• Preserve timestamps.
• Preserve uncertainty.
• Preserve approval status.
• Preserve conflicting evidence.

Bad summaries create confident agents with missing facts.

Context changes tool use#

The context layer should help the agent decide what to do next:

• Enough evidence to answer.
• Need another tool call.
• Need approval.
• Need a stronger model.
• Need to stop.

That is why context engineering belongs in the workflow, not just the retrieval backend.

Build it in Codelit#

Try this:

Design context engineering for an agentic internal support system. Include source routing, scoped retrieval, freshness, memory, tool results, compression, citations, approval-aware context, and evals.

Map the context architecture

Better prompts help. Better context changes the system.