TPipe - LLM Pipeline Framework for AI Agents

TPipe is an Agent Operating Substrate - infrastructure that autonomous AI agents run inside, rather than a library they call. It's built for agents that operate for days, not minutes, and it addresses the two failure modes that most agent frameworks leave to the developer: memory loss across long sessions and silent failure when something goes wrong mid-task. Core capabilities:
- Reasoning Pipes - structured JSON control over token prediction that forces any LLM to think in a defined pattern (Chain-of-Draft, Role Play, and 6 other modes), independent of whether the model has native reasoning. Reported 75% token reduction and 78% latency decrease with Chain-of-Draft.
- ContextBank - thread-safe, persistent memory across distributed runs, with weighted lorebook-style retrieval, substring-triggered activation, and token-budget-aware selection. Designed to survive 120+ turn conversations without degradation.
- Pipelines - sequential orchestration with declarative pause, resume, and jump points, plus 18 named hooks (DITL) across the pipe/pipeline lifecycle for validation, transformation, and human-in-the-loop review.
- Manifold, Junction, and DistributionGrid - three distinct multi-agent patterns: manager-worker state machines, democratic voting/handoff between agents, and peer-to-peer cluster orchestration (8,773 LOC of distributed infrastructure).
- PCP (Pipe Context Protocol) - secure, multi-language function/tool calling across Stdio, HTTP, Python, Kotlin, and JavaScript transports, with per-language access controls.
- KillSwitch and loop limits - forced, non-catchable termination on token overrun or runaway loops, so failures stop cleanly instead of degrading silently. Deployment: TPipe is JVM-native (Kotlin, JVM 24) and headless-first. It also ships as a GraalVM Native Image - a ~50MB binary with no JVM at runtime, sub-128MB memory footprint, millisecond startup, and support for Linux, macOS, Windows, and ARM. It works with any LLM reachable through a standard transport executor, including AWS Bedrock, Ollama, and OpenRouter.
Proof points from production use: Autogenesis, a headless game-master agent, runs 25 rounds and 120+ turns of inference without degradation. TPipeWriter maintains coherence across a 300-page manuscript through multi-stage refinement. TStep is a multi-agent debugger that tests, captures crashes, and steps through code via DAP/ADB. TPipe is a persistent substrate with topdown token governance and cross-session memory built in from the start.