The open-source AI platform: the harness is the platform, not the model
An open-source AI platform is built on the harness, not the model: the orchestration layer that turns a model into an agent. Open harnesses make that layer fully controllable, and routing sends 90 percent of work to cheap open-weight models while reserving frontier models for the ten percent that needs them. Pexon builds these architectures.
The model race is over. The platform question has replaced it.
The last year produced a large, cheap open-weight model that reset the conversation: many models are now inexpensive, solid and sometimes runnable locally, which has collapsed the entry barrier to AI. But between a model and a usable AI worker sits a layer the public debate mostly skips — the agent harness.
A harness is the scaffold that turns a model into an agent. It owns the prompting rules and system context, the tool calls, the agentic loop of plan-execute-check-iterate, session and context management, the sandboxes and storage, and the user interface. The model is the engine. The harness decides what the engine may do.
That is why the platform question has moved. It is no longer 'which model' but 'which harness' — and whether you control the layer between the model and your work.
The control claim
Closed harness vs open harness, layer by layer
| Layer | Closed harness | Open harness |
|---|---|---|
| Prompting rules | Not changeable | Fully adaptable |
| Agentic loop | Hard-wired | Swappable as a plugin |
| Tools and skills | Partly extensible | Everything is a plugin |
| Sessions, storage, sandbox | Vendor decides | Under your control |
| UI | Fixed | Adaptable |
| Model choice | Restricted | Any: frontier, open-weight, local |
| Fixing a limitation | Vendor feature request | Self-solvable via creator mode |
| Transparency | Black box | Session logs, trajectory, traceable |
A closed harness is a car you may swap the engine in, but the seat, steering and chassis stay fixed. An open harness lets you rebuild the whole vehicle. That difference — not the licence — is what makes the platform open.
Everything is a plugin: the property that makes the platform yours
The decisive phrase in open harnesses is 'everything is a plugin'. That does not mean skills and MCP servers — it means every capability: tools, skills, sessions, sandboxes, storage, even the loop and the UI. When you hit a boundary in a closed platform, the only path is a feature request to the vendor or a workaround. In an open one, you tell the agent to switch to creator mode and build the fix as a plugin.
The platform extends itself, and the knowledge stays in the company. Every solved problem can become a skill or a plugin, so sessions stop being throwaway work and start compounding into an institutional asset.
That self-extending property is the practical difference from 'open source as a licence'. It is why the architecture is the deliverable, not the tool.
Executable artefact
Creator mode: the limitation becomes a plugin
Where a closed platform would mean filing a vendor request, an open harness means instructing the agent to build the capability itself. The trajectory log and session export are what make that reproducible.
# Open a session, hit a limitation, extend the platform in the same run.
opencode run "Switch to creator mode. Build a plugin that
adds <capability> to this harness, wire it as a tool, and
document it. Then use it on the current task." --thinking
# The session is not throwaway — it becomes the next skill.
opencode session export --session <id> > docs/skills/<capability>.mdThe self-extending loop: a session becomes a skill, a skill becomes a plugin, and the platform grows without a vendor in the path.
The routing principle that makes open source the cost lever
- Route 90 percent of the work to a good cheap model. Standard work, documents, email, research and every automated agent run on an open-weight model with routing. The bulk of the workload does not need a frontier model.
- Reserve the frontier model for the ten percent that earns it. Development, complex analysis, strategic work and power users get the expensive model where output quality is the product. Ten percent, not one hundred.
- Measure cost-to-completion, not token price. A harness that finishes a task in a fifth of the time is cheaper even at the same token price. Efficiency and iteration count decide the bill more than the rate card.
- Keep the platform harness-agnostic. Model, harness and components stay swappable, so a price change or a better model never forces a re-platform. Independence is the point of the architecture.
For 900 of 1,000 employees, a good open-weight model with routing is enough. The 100 who need a frontier model are the reason the platform must be able to reach one — without being locked to it.
The economics: efficiency beats the price card
The strongest argument for the open architecture is model routing, and the numbers follow from the same principle as the manager-worker split elsewhere on this site. A company of 1,000: 100 power users work with frontier models where maximum quality is justified. The other 900 — and every automated agent — use the right model via routing, usually a cheap but solid open-weight one.
The saving compounds because the expensive tokens are confined to the cases that genuinely need them, and because agents that run around the clock run on the most economical models. Scaling no longer means a price shock.
The honest caveat is readiness. Many open harnesses still feel like a developer preview: bugs, context regressions in long sessions, occasional UI issues. That is normal for young open-source projects and it is improving fast — which is exactly why professional engineering is part of the architecture rather than optional. A company should not install an open harness and hope; it should build a platform with clear quality and operating standards.
Open source wins on control and cost, not because it is free. The licence is a side effect; the control is the point.
The honest risks. Open-source harnesses are powerful but not yet 'just works': bugs and long-session context problems need professional engineering. Cheap models have blind spots — a model without visual perception fails at design work that needs to see, screenshot and iterate, so model choice per task is essential. Skills are not portable across harnesses: changing harness means reworking them. Third-party plugins need a security review before installation. And an open harness can seem over-confident — results must be checked for the purpose they serve.
Keep reading
- What is LiteLLM? The one API in front of every LLM provider — the gateway layer this architecture routes through
- LiteLLM vs OpenRouter — the gateway decision for multi-model apps — self-hosted gateway against hosted routing
- OpenCode: the open-source coding agent challenging Claude Code — one concrete open harness, with custom agents
- open weight model hosting on hardware you control — the money page for the open-weight side of the routing rule
- GPU inference sizing: matching hardware to a token budget — the cost model once the checkpoint is chosen
- Private AI hub: open weights on hardware you control — the cluster this post belongs to
Questions we get asked about open-source AI architecture
What is a harness in AI?
A harness is the orchestration layer that turns a model into an agent: prompting rules and system context, tool calls, the agentic loop of plan-execute-check-iterate, session and context management, sandboxes and storage, and the user interface. The model is the engine; the harness decides what the engine can do. Open-source harnesses make that layer fully controllable and swappable.
Why is the harness the platform and not the model?
Because the harness controls everything the model can do — tools, loops, sessions, UI — and because it is what makes the model exchangeable. A closed harness locks the whole vehicle; an open one lets you rebuild it. That control is why the platform question has moved from which model to which harness.
Can open source really save money without losing quality?
Yes, through routing: send the bulk of work to a good cheap open-weight model and reserve the frontier model for the ten percent that needs it. The cost lever is efficiency and routing, not downgrading. A harness that finishes a task in a fifth of the time is cheaper even at the same token price.
Are open-source harnesses production-ready?
The honest answer is 'almost'. Many feel like a developer preview: bugs, context regressions in long sessions, occasional UI issues. With professional engineering and clear operating standards they are ready for production — which is exactly why companies run them as a platform rather than installing them and hoping.
Next step
Get the open platform decision out of the slide deck
Two weeks, fixed price. We map which layers of your AI stack a closed vendor actually controls, sketch the open harness and routing design that returns them to you, and hand over the build plan with the cost comparison. The plan is yours whether or not we build it.