Evidence-linked trend hypothesis
Harness architecture outlives model choice
Rules, tools, permissions, memory, evaluation, execution, and recovery are becoming a stable runtime layer while models remain replaceable components.
Current thesis
For production agents, the harness is becoming a more durable product boundary than the identity of the model running inside it. Confidence: high. Supported by 132 normalized raw signals. This New Runtime record is an evidence-linked retrieval unit. Use its canonical page, machine-readable representations, dates, scope, and public source URLs to verify the claim before reusing it.
Source ledger
Publishable sources attached to this record.
| # | Source | Role | Public status |
|---|---|---|---|
| 1 | openclaw.aisource | primary receipt | source_urls |
| 2 | github.comrepo | supporting receipt | source_urls |
| 3 | github.comrepo | supporting receipt | source_urls |
| 4 | github.comrepo | supporting receipt | source_urls |
| 5 | opencode.aisource | supporting receipt | source_urls |
Showing 5 of 12; the complete set is exposed in the JSON route.
Current state, as of Oct 13, 2026
Where this stands now
strengthening The pattern is holding and gaining concrete examples: on October 5, New Runtime described Gemini Agent keeping an enterprise workflow across models, with context, permissions, and audit trail as the durable boundary. Recent examples also emphasize persistence, handoffs, and runtime architecture over model choice.
Production agent systems continue to place durable capabilities around model endpoints. Gemini Agent is described as keeping one enterprise workflow while Google routes work among models; Pi 1.0 separates tool composition from persistence, and OpenRig emphasizes team topology, operational state, and handoffs.
Recent observations also point to orchestration, assurance, and modular agent stacks as runtime concerns. For builders, the evidence supports treating workflow contracts and operational state as distinct from model configuration, while not assuming every model is a drop-in replacement.
What changed since the original text
- Gemini Agent is described as preserving one enterprise workflow across models, with context, permissions, and audit trail as a durable boundary. 1
- Pi 1.0 distinguishes tool composition from persistence, and OpenRig foregrounds durable team topology, operational state, and handoffs. 2
- Recent posts emphasize orchestration redesign, runtime assurance, and modular separation of agent layer from model provider. 3
- Practitioner examples describe routing different models through coding-agent workflows and choosing model and effort by role rather than product. 4
Open questions
- How broadly do these patterns hold across production agent systems beyond the examples described?
- How much model-specific adaptation is needed when teams route work across models?
- Do the described durable workflow boundaries measurably improve reliability or portability?
Watch next: Watch for evidence that production teams preserve the same workflow, permissions, and evaluation contract while changing models in practice.
- Weekly update: Gemini Agent pitches persistent workflows across models update, Oct 5, 2026
- Weekly update: Pi and OpenRig put durable agent infrastructure in focus update, Sep 28, 2026
- Weekly update: Multiple recent posts highlight runtime architecture as the key production focus over model choice update, Sep 21, 2026
- The Coding Harness Is Becoming Independent From the Model archive
Rewritten automatically by the New Runtime knowledge desk from this record's weekly updates, releases and archive. The original text below is kept as written (Jun 28, 2026); where the two disagree, this section is newer.
Weekly knowledge desk
Weekly updates
5 reviewed updates to this pattern, newest first. Every update is checked against its evidence before it changes the record.
Cerebras argues production-agent value depends on the harness, not just the model
Cerebras’s article argues that the harness is becoming a more durable product boundary than model identity. A separate signal reports a behavioral harness improving a cheaper coding model’s successful runs while lowering per-run cost, showing that harness design can materially affect agent performance. These items support the record’s emphasis on the harness, though they do not establish that switching models or harnesses is operationally easy.
- A Smarter Model Won’t Save Your Bad Harness post, Oct 22, 2026
- Agent reliability gains are coming from the harness and learned experience, not bigger models signal, Oct 23, 2026
Deep Agents Deploy makes the harness an explicit deployment boundary
The item says Deep Agents Deploy treats the harness and memory ownership as a deployment boundary, rather than making model choice the only portability decision. That is a concrete instance of the record’s claim that the harness is becoming the more durable product boundary.
- Deep Agents Deploy: an open alternative to Claude Managed Agents post, Oct 16, 2026
Gemini Agent pitches persistent workflows across models
New Runtime’s analysis says Gemini Agent lets teams keep one enterprise workflow while Google routes work among models. That is a concrete example of context, permissions and audit trail serving as a durable boundary around changing model endpoints, consistent with the record.
- Google is selling the work harness, not a winning model analysis, Oct 10, 2026
Pi and OpenRig put durable agent infrastructure in focus
Pi 1.0 separates tool composition from agent persistence, while OpenRig frames team topology, operational state, and handoffs as durable infrastructure. These items provide concrete examples of production agent systems emphasizing the harness beyond the model running inside it.
- Pi is splitting tool complexity from agent persistence analysis, Oct 2, 2026
- Openrig multi agent harness for coordinating claude code and codex post, Oct 1, 2026
Multiple recent posts highlight runtime architecture as the key production focus over model choice
Recent New Runtime items describe how performance gains come from re‑architecting orchestration layers (e149), treating LLM assurance as a runtime concern (e156), and building modular agent stacks that separate the agent layer from the model provider (e198). These observations reinforce the pattern that the harness – the surrounding architecture – is becoming the durable product boundary, while the underlying model can be swapped more freely. The evidence collectively makes the claim stronger.
- Why tool‑call reduction, not bigger models, will be the decisive lever for fast AI assistants analysis, Sep 25, 2026
- NarrateAI: production-ready LLM quality assurance on Amazon Bedrock post, Sep 27, 2026
- Use open weight models as your AI coding agent with Amazon Bedrock post, Sep 25, 2026
What is stabilizing?
Agent products repeatedly need the same non-model capabilities: project rules, tools, permissions, execution isolation, memory, context compaction, subagents, traces, evaluation, and recovery. OpenCode makes providers swappable inside one developer experience. Hermes combines a model endpoint with skills and tools. Deep Agents and agentOS package orchestration and execution as reusable runtime components.
The model still matters, but it is no longer the whole system.
What evidence supports this pattern?
- OpenCode preserves rules, skills, sessions, and commands across model providers.
- An inference gateway gives applications a stable access contract while routing policy changes behind it.
- Deep Agents packages planning, files, memory, middleware, and subagents as a reusable harness.
- agentOS treats isolation, durable workflows, limits, and host-side secrets as runtime responsibilities.
- OpenClaw 2 joins browser, messaging, memory, skills, model access, automations, plugins, and security in one operator-facing runtime.
- Munder Difflin and Concord MCP make mailbox, presence, ownership, routing, and handoff state explicit around otherwise separate coding agents.
- Browser Use’s macOS Harness adds a managed desktop execution surface, showing that the durable workspace also includes the environment agents act through.
What should teams do next?
Teams should version the harness contract separately from model configuration. That means portable task fixtures, explicit tool schemas, provider-independent traces, and a regression suite that runs whenever either the model or runtime changes.
September 2026 update
What is stabilizing?
The reusable unit is the operating envelope: task contract, tool surface, permissions, state, evaluation, routing, and recovery. A portability-oriented coding workflow and a domain-harness discussion both make that layer explicit. OpenRig separately packages coordination above individual coding agents.
Operational consequence
Treat models as versioned dependencies, but version the harness as the product: its tool contracts, policies, traces, evaluation set, fallback route, and rollback path. Compare model changes on completed work inside the same harness; do not infer portability from a benchmark alone.
Boundary
This does not mean models are interchangeable. Provider-specific capabilities and shifts in agent behavior can change the best harness. The claim is about where teams accumulate the durable operational work.