Enterprise Implementation Playbook

Executive Summary

Transitioning an enterprise engineering department from consumer-grade monolithic AI plugins to a production-grade, multi-agent AI software factory requires precise tactical execution.

This playbook provides a step-by-step implementation roadmap synthesized from benchmark evaluations across four core topologies. It details how to configure open agent harnesses (OpenCode via Agent Client Protocol / ACP), establish model routing layers (OpenRouter), set up local inference servers (LM Studio), and harden subagent specifications to achieve up to 100% data sovereignty and 80%+ cost reduction.

1. Hardware Sizing & Infrastructure Matrix

2. Environment Setup Runbook

Step 1: OpenCode Harness Setup

OpenCode is an open-source CLI harness operating over the Agent Client Protocol (ACP), decoupling agent UI from underlying inference APIs.

1. Deploy the OpenCode CLI binary to team developer machines.
2. Verify ACP engine status with opencode version.

Step 2: Configure Model Routing

Set configuration in global provider config files to manage cloud routers and local endpoints.

Step 3: LM Studio Setup

1. Launch LM Studio on developer workstations.
2. Load models: Gemma 4 12B IT (8-bit) and Qwen 3.5 27B Instruct (Q4).
3. Enable local server on port 1234 with context window size set to 40,960 tokens.

3. Hardening Subagent Configurations

Subagents operate as automated micro-executors. To prevent unexpected behavior, small-model diff corruption, and runaway token consumption, create explicit subagent specs in repository roots:

Implementer Configuration Specification

• Role: Bounded local implementation specialist.
• Tools Enabled: write (force whole-file overwrites), edit (disabled to prevent search-and-replace diff errors), bash (permitted for test execution), ls (disabled), find (disabled).
• Execution Ceilings: max_steps set to 12 (hard termination limit).
• Strict Execution Rules: Execute only the task defined in the Handoff Packet. Never scan virtual environment or git dependency directories. Write complete, working files using whole-file write tool. Run unit tests before completing.

4. Spec-Driven Handoff Execution Workflow

To demonstrate operational execution, we trace the build of a standardized enterprise service:

Handoff Packet Structure

Passed directly from the Architect agent to the Implementer subagent:

1. Objective: Itemized target feature (e.g., POST endpoint implementation).
2. Constraints: Approved frameworks (FastAPI, Pydantic v2) and mandatory file targets (app/main.py, tests/test_complaints.py).
3. Verification: Automated test command specification (pytest tests/test_complaints.py).

5. Operational Realities & Resolution Protocols (Tips & Tricks)

During enterprise rollouts, engineering teams will encounter specific behaviors across small models and local inference backends. Use these battle-tested resolution protocols:

Operational Nuance 1: Small Model Diff Handling

• Symptom: Models with <15B parameters hallucinate line numbers or fail to match exact string patterns when using search-and-replace edit tools.
• Resolution: Disable edit tool and enable write tool in subagent configuration. Force the subagent to re-write entire files. For small files (<300 lines), whole-file writing is 100% deterministic and eliminates syntax corruption.

Operational Nuance 2: Execution Loops & Step Ceilings

• Symptom: A subagent encounters a failing test and repeatedly executes tests without fixing the bug, consuming compute or hanging local hardware.
• Resolution: Enforce step limits (max_steps: 12) in subagent config. Force the subagent to report failure back to the primary Architect agent after 12 steps, allowing the higher-capacity Architect model to diagnose the issue.

Operational Nuance 3: Context Degradation ("Context Rot") from Virtual Environments

• Symptom: Harness context token count explodes to 200k+ tokens on step 2 due to recursive file searching.
• Resolution: Add .venv, node_modules, .git, and dist to ignore files. Instruct subagents explicitly in prompt instructions to ignore dependency directories.

Operational Nuance 4: Local Server Timeouts in LM Studio

• Symptom: Local inference calls fail with connection timeout during heavy subagent execution.

• Resolution: In LM Studio server settings:

  1. Increase Keep Model in Memory Timeout to Infinity.
  2. Increase Context Window Size explicitly to 40960.
  3. Enable GPU Acceleration offloading to 100% of layers.

6. Enterprise Migration Roadmap: A 4-Phase Execution Blueprint

1. Phase 1: Protocol Standardization (Weeks 1–2): Replace monolithic plugins with open ACP harnesses (OpenCode). Establish central OpenRouter billing with Zero-Data-Retention (ZDR).
2. Phase 2: Spec-Driven Integration (Weeks 3–4): Require all engineering teams to embed Project Constitutions (mission.md, tech-stack.md) and hardened agent YAMLs into repository templates.
3. Phase 3: Hybrid Local Offloading (Weeks 5–6): Deploy 16GB local developer workstations with LM Studio. Route subagent implementer tasks locally while keeping high-level cloud planning. Achieve 80%+ metered cloud cost savings.
4. Phase 4: Full Sovereign Infrastructure (Weeks 7–8): Deploy 32GB/64GB local rigs running local Qwen 27B orchestrators for classified, air-gapped repositories. Achieve complete IP privacy and $0.00 metered compute costs.

Artifact Reference: 2552e8f0-27c9-4d70-8681-634429b9e6fa