HarnessFactory is a unified codebase and execution foundation that re-implements 13 representative contemporary agentic scaffolds under a shared four-module protocol $\Pi$ and a common runtime kernel. By formalizing diverse agent scaffolds—including ReAct, Plan-and-Execute, ReSum, Flash-Searcher, GAM, MemoBrain, AggAgent, OAgent, AgentFold, HiAgent, DeepAgent, ROMA, and AOrchestra—into modular four-tuples of memory, planning, action, and capability orchestration $(M, P, A, F)$, HarnessFactory validates the expressiveness of the protocol space $\mathcal{H}_\Pi$ and provides structured, protocol-compliant source material for meta-harness synthesis.

HarnessFactory

The protocol-compatible harness seed bank $\mathcal{B}_0$ is an initial catalog of $K_0 = 13$ hand-written agent harnesses instantiating the four-module protocol $\Pi$. The bank serves a dual operational role in just-in-time harness generation:

1. **Synthesis Anchoring**: Subsets of scaffolds sampled from $\mathcal{B}_0$ provide in-context reference examples to condition Stage-I task-adapted harness synthesis by a teacher model.
2. **Evolutionary Incumbent Archive**: As tasks are executed, $\mathcal{B}_0$ establishes the initial population of prior designs against which newly proposed candidates are benchmarked, progressively expanding into $\mathcal{B}_n \supseteq \mathcal{B}_0$ where retained harnesses are associated with their task identifier, observed reward, latency, and monetary cost.

Protocol-Compatible Harness Seed Bank

The four-module protocol $\Pi$ maps heterogeneous agent architectures into comparable coordinates within the product space $\mathfrak{M} \times \mathfrak{P} \times \mathfrak{A} \times \mathfrak{F}$, demonstrating the expressive breadth of the modular design space:

- **Canonical ReAct**: Formalized as $h_{\text{ReAct}} = (M_{\text{full}}, P_{\emptyset}, A_{\text{react}}, F_{\text{all}})$, retaining complete unmanaged history ($M_{\text{full}}$), omitting an explicit planner via a null directive ($P_{\emptyset}$), utilizing the standard reasoning-action loop ($A_{\text{react}}$), and exposing the unrestricted tool/skill registry ($F_{\text{all}}$).
- **Engineered ReAct Variants**: Formalized as $(M_{\text{compact}}, P_{\text{todo}}, A_{\text{react}}, F_{\text{all}})$, where execution retains a ReAct-style action kernel ($A_{\text{react}}$), but incorporates working-memory compression near the context limit ($M_{\text{compact}}$) and maintains an explicit task agenda ($P_{\text{todo}}$).
- **Recursive Agent Architectures**: Formalized as $h_{\text{rec}} = (M_{\text{subproblem}}, P_{\text{decomp}}, A_{\text{rec}}, F_{\text{route}})$, where the orchestrator decomposes subproblems into a todo agenda ($P_{\text{decomp}}$), spawns subagents via recursive execution ($A_{\text{rec}}$), maintains isolated memory contexts for each subagent ($M_{\text{subproblem}}$), and routes scoped tool/skill subsets to subagents ($F_{\text{route}}$).

Four-Module Factorization of Canonical and Recursive Agent Scaffolds

Model-as-a-Harness is an agent design paradigm in which an agentic operational scaffold is formalized as a machine-generatable artifact synthesized dynamically by a trained meta-agent. Rather than deploying foundation models within manually engineered, fixed harnesses, Model-as-a-Harness treats harness synthesis as a learned capability: the meta-agent generates, validates, and refines a customized four-module harness (governing memory, planning, action, and capability orchestration) at inference time, under which an arbitrary off-the-shelf foundation model executes.

Model-as-a-Harness

In program-based harness generation, candidate harnesses are evaluated within nested formal spaces:

$$\mathcal{G} \supseteq \mathcal{H}^{\text{syn}} \supseteq \mathcal{H}_\Pi \supseteq \mathcal{H}_\Pi^{\text{exec}}$$

1. **Raw Generation Space ($\mathcal{G}$)**: The unrestricted output space of programs emitted by a generative model.
2. **Syntactically Valid Harnesses ($\mathcal{H}^{\text{syn}}$)**: The subset of programs that parse cleanly without syntax errors.
3. **Protocol-Compliant Harnesses ($\mathcal{H}_\Pi$)**: Programs adhering strictly to the shared protocol $\Pi$, which specifies modular schemas, lifecycle interfaces, validation rules, and execution semantics for $(M, P, A, F)$.
4. **Executable Harness Subset ($\mathcal{H}_\Pi^{\text{exec}}$)**: The set of protocol-compliant harnesses that run successfully without runtime exceptions, tool-call failures, or interface crashes under an environment kernel.

Harness Protocol Spaces and Syntactic Subsets

University of Michigan - Ann Arbor

Ch.3 Harness Synthesis and Factory Patterns - Autonomous Agent Control Planes: State Scaffolding and Resilient Execution @ University of Michigan - Ann Arbor

Ahead-of-Time versus Just-in-Time Harness Paradigms - Autonomous Agent Control Planes: State Scaffolding and Resilient Execution @ University of Michigan - Ann Arbor

HarnessFactory and Protocol-Compatible Seed Libraries - Autonomous Agent Control Planes: State Scaffolding and Resilient Execution @ University of Michigan - Ann Arbor

Learn Before

Related

Learn After