Context policies

ContextPolicy controls what prior information an agent sees on each step: earlier step outputs, the original run input, and size limits. It applies in both Linear workflows and Graph workflows. Graph state from node outputs may also arrive via exec_config.initial_state and the runtime map.

Agent overview: Defining agents. workflow specification types: Workflow specification.

ContextPolicy fields

{
 "context": {
 "include_prior_steps": "last",
 "include_run_input": true,
 "max_prior_step_chars": 4096
 }
}

Engine default for max_prior_step_chars is 4096 (verified in types.rs).

include_prior_steps modes

"last" (default)

Only the immediately preceding step output is included. Typical for pipelines where each stage refines the previous result.

{
 "context": {
 "include_prior_steps": "last",
 "include_run_input": true,
 "max_prior_step_chars": 4096
 }
}

"all"

All completed prior steps are included up to max_prior_step_chars budget (implementation applies per-step truncation rules). Use when downstream agents need full chain context; watch token costs.

"none"

Prior step text is omitted. The agent still receives run input if include_run_input is true. Useful for isolated classification steps that should not be biased by earlier prose.

{
 "context": {
 "include_prior_steps": "none",
 "include_run_input": true,
 "max_prior_step_chars": 4096
 }
}

include_run_input

When true, the original input passed to workflow.run(input) or POST /v1/runs is available to the agent prompt assembly. Set false when later steps should only see structured prior outputs or memory, not the raw user message.

Example for a summarize-only second step:

{
 "context": {
 "include_prior_steps": "last",
 "include_run_input": false,
 "max_prior_step_chars": 8192
 }
}

max_prior_step_chars

Prevents unbounded prompt growth from large intermediate outputs. Prior text is trimmed before provider request (ProviderRequestSent records prompt_size_bytes in trace, not content).

Raise the limit for long-document workflows:

{
 "context": {
 "include_prior_steps": "all",
 "include_run_input": true,
 "max_prior_step_chars": 16384
 }
}

Balance against model context window and cost.

Graph workflows and initial_state

Graph nodes may write keys via outputs:

{
 "id": "n-classify",
 "step_ref": "s-classify",
 "outputs": { "route": "category" }
}

Seed or resume graph state through exec_config:

{
 "exec_config": {
 "initial_state": {
 "route": "billing",
 "observation": "Customer mentioned invoice 8842"
 }
 }
}

Context policy and graph state compose: policy controls step output text; graph state supplies structured fields the engine merges into agent context per scheduler rules.

Memory vs context policy

Use shared memory for multi-agent handoff that must survive truncation limits:

{
 "memory_config": {
 "memory_type": "shared",
 "scope": "workflow",
 "namespace": "pipeline-state"
 }
}

Multi-agent linear example

Step 1 researcher with full user input; step 2 writer with last step only:

{
 "agents": [
 {
 "id": "a-research",
 "name": "researcher",
 "role": "Research",
 "instructions": "Research the topic.",
 "context": {
 "include_prior_steps": "none",
 "include_run_input": true,
 "max_prior_step_chars": 4096
 }
 },
 {
 "id": "a-write",
 "name": "writer",
 "role": "Writer",
 "instructions": "Write a summary of the research.",
 "context": {
 "include_prior_steps": "last",
 "include_run_input": false,
 "max_prior_step_chars": 8192
 }
 }
 ]
}

Trace visibility

Context assembly is not a separate trace event. Related events:

• AgentInvoked with input_size_bytes
• ProviderRequestSent with prompt_size_bytes

No prompt text appears in traces (trace data policy).

Related pages

• Defining agents
• Linear workflows
• Graph workflows
• Trace data policy