Extract the reusable execution core
The agent loop consumes a tool-runtime contract. Model calls and run events stay in the library, while product toolsets and permissions remain in the host application.
Knolume Runtime.
Composable execution primitives for AI applications
A public TypeScript library for agent runs, tools, context, memory, retrieval, guardrails and sub-agents. Host applications provide models, tools, identity and product policy.
- Area
- Agent systems
- Implementation focus
- Agent loop, tool pipeline, context, memory, retrieval, guardrails and protocol contracts.
- What you can inspect
- Inspect public source, a deterministic example and the versioned v0.4.0 release.
The problem
An agent application needs more than a model loop. Tool effects, approvals, context limits, cancellation and records must be composed without forcing every product to inherit the same business rules or entry point.
How the pieces fit
Composable agent runtime
Host model & policy
Agent loop & context
Tool execution & approvals
Guardrails · traces · recovery
Public library · Product decisions stay in the host
The library exposes runtime modules and protocol contracts. Applications supply their own models, prompts, concrete tools, authorization, storage and transport; the diagram is not a standalone hosted service.
Key engineering decisions
Gate effects before they escape
Guarded execution can buffer generated events while input checks run. Tool execution waits for the input-effect gate; a rejected run discards buffered output instead of presenting it as completed work.
Publish an integrable boundary
A versioned package asset, public entry points and a deterministic example let another application inspect and integrate the library without assuming it contains a CLI or product server.
Reliability & failure handling
Cancellation and recovery are explicit
Run cancellation propagates through execution. Session, trace and guardrail code distinguish interrupted or rejected work from success.
Host policy is not smuggled into the library
The host owns identity, tenant isolation, credentials, approvals and concrete tools. Runtime guardrails do not replace those authorization boundaries.
Offline checks have a bounded claim
The repository provides tests, coverage and a deterministic example. Those checks do not establish live-provider quality or production operation for a consuming product.
Evidence & scope
Source links are pinned to the reviewed revision so you can follow each implementation detail.
Agent loop
Run events, model steps, cancellation and tool-runtime composition.
src/harness/agent/loop.tsTool execution boundary
Argument validation, permissions, approvals and outcomes.
src/tools/execution-pipeline.tsGuardrail runner
Buffered checks, effect gating and recovery handling.
src/harness/guardrails/runner.tsDeterministic example
A runnable model-free integration example.
examples/basic-agent.mjsVerification
Inspect behavioral tests; this link does not assert the latest CI run passed.
src/harness/agent/loop.spec.tsWhat this case establishes
View the archived predecessorPublic source and a v0.4.0 GitHub Release are available. This is a library, not a hosted application or a drop-in replacement for the archived coding-agent application.