Overview
TrueForge is an open-source agent harness that surrounds model calls with a reusable agent contract, durable sessions and turns, context controls, mediated tools, sandboxes, human checkpoints, and client and deployment surfaces. The series moves from runtime boundaries to execution lifecycle, context budgeting, bounded side effects, and local-to-hosted operation, then weighs benchmark evidence and current limits. Its value is making information, authority, intervention, and state explicit; operators still own providers, identities, external correctness, retries, and workload-specific evaluation.
Learning outcomes
- TrueForge makes context, authority, human intervention, and state explicit around model-driven work.
- Agent definitions separate reusable behavior from session history, while turns provide bounded, streamable, pausable execution.
- Deferred tools, progressive skills, subagents, offloading, Code Mode, and compaction address different context pressures with distinct costs.
- In the project’s reported 14-task benchmark, TrueForge’s Opus configuration matched Claude Managed Agents’ mean solved-task count with fewer reported tokens and dollars; this is not universal proof.
- Adoption remains conditional: operators must secure providers and deployment, review skills and annotations, handle idempotency, and test their own workload.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
What does TrueForge add around a model call?
TrueForge adds an execution loop that routes tools, manages context, pauses for human input, and persists sessions and turns; it does not own provider or external-system correctness.
How does TrueForge keep long-running agent work manageable?
It manages input and runtime context with deferred tools, progressive skills, subagents, offloading, Code Mode, and compaction, trading context savings for latency, coordination, sandbox use, or fidelity.
How does TrueForge handle approvals, OAuth, and clarifying questions?
It represents each as a typed required action that can end a turn and resume through new input; approval grants permission, OAuth authorizes access, and questions supply missing information.
Does TrueForge’s benchmark prove it is better than other agent runtimes?
No. In the project’s three-trial, 14-task comparison, one TrueForge Opus configuration matched Claude Managed Agents’ mean solved-task count with fewer reported tokens and dollars, but the result is workload-specific and does not establish universal superiority.