Models and agent execution
Providers
VOCION_LLM_PROVIDER picks the default: anthropic (default), openai or bedrock. An agent overrides it with harness.modelProvider and model; a workspace sets defaults.model.
| Provider | Credentials | Default main model |
|---|---|---|
| Anthropic | ANTHROPIC_API_KEY | claude-sonnet-5-5 |
| OpenAI | OPENAI_API_KEY | — |
| Bedrock | the workspace's stored AWS key, else the AWS SDK chain (AWS_BEARER_TOKEN_BEDROCK, access keys, profile, instance role); AWS_REGION | us.anthropic.claude-sonnet-4-6 |
Per role overrides: VOCION_LLM_PROVIDER_<ROLE> and VOCION_LLM_MODEL_<ROLE> for MAIN, CLASSIFIER, EMBEDDER, SKILLTURN, EXTRACTOR. A workspace that stores its own vendor key under Developers is billed on its own account.
On AWS with vocion-stack, the host role can already invoke Anthropic models on Bedrock (in-region and through us. inference profiles).
Embeddings
OpenAI text-embedding-3-small by default; Bedrock Titan (amazon.titan-embed-text-v1) when VOCION_EMBEDDING_PROVIDER=bedrock or defaults.embeddingProvider: bedrock. Both produce 1536 dimensions, the width the database stores; switching providers means re-embedding. Rerank uses VOCION_RERANK_MODEL.
Where a turn runs
harness.runsOn on an agent:
| Value | Runs |
|---|---|
in-process | In the app's own process (default unless the provider is Bedrock) |
agentcore-container | In Vocion's agent-runtime container on Amazon Bedrock AgentCore Runtime (default for modelProvider: bedrock) |
aws-managed-harness | Handed to AWS's managed agent harness |
external-worker | A long-running worker outside the app (see Runners) |
The first two behave the same way: same tools, same trust rules, same record. Tools called from AgentCore come back to the app at VOCION_TOOL_ENDPOINT_URL, which must be reachable from AWS.
AgentCore setup
infra/agentcore/provision.sh— ECR, roles, AgentCore Memory; writes outputs to SSM under/vocion/agentcore/<env>/.infra/agentcore/deploy-runtime.sh— builds and deploys the runtime container (arm64). This is a second deploy, separate from the app's.- Set
VOCION_AGENT_RUNTIME_ARN,VOCION_AGENTCORE_REGION(defaultus-west-2),VOCION_TOOL_ENDPOINT_URL, and optionallyVOCION_AGENTCORE_MEMORY_ID. With vocion-stack,agentcore_enabled = truegrants the hostInvokeAgentRuntimeandInvokeAgentRuntimeForUser.
Kill switches: VOCION_DISABLE_AGENTCORE=1, VOCION_DISABLE_RUNTIME=1 (force in-process), VOCION_DEFAULT_RUNS_ON.
Prompt caching
On by default, nothing to configure. Turn off per agent (harness.promptCache) or process-wide with VOCION_PROMPT_CACHE=0.