Repository navigation
Is something like OpenRig only relevant for frontier models until ~2028? #2298
Replies: 2 comments 1 reply
|
Thanks, @porterchild, for trying it out and for a thoughtful question, and sorry this sat unanswered for several hours. Some of what OpenRig does doesn't depend on how capable the model is. It runs several agents as separate sessions in separate terminals, and a single session, however capable, ends or runs out of context. So the work itself lives outside any one session: who's on the team and how to reach them, a durable queue of work that survives a session ending, handoffs that record who holds what, and status a person can read without asking each agent. Those are about several agents working over time, more than about correcting for a weak model. It also isn't built around one model. You choose each agent's harness (Claude Code, Codex, Pi, or a community harness) and its model separately, so a smaller or cheaper model can take a role where it's good enough, next to a frontier model where it isn't. If you try smaller or local models in a team, we'd be glad to hear what you find here, including where the structure helps them and where it just costs tokens. Your idea of different modes by model capability is a useful way to frame that. |
|
I believe the core question isn’t actually about the models themselves, but about the harnessing clients attached to them (e.g., OpenAI and Claude desktop/CLI applications). As these first-party clients get significantly better at handling sub-agents, sub-processes, and nested workflows, it's a completely valid question to ask: for typical, single-stream development loops, a tool like OpenRig may quickly become overkill. In the near term, however, OpenRig-class harnessing will stick around because it solves a fundamentally different and more complex operational problem: Vendor Neutrality Over Single-Ecosystem Focus: First-party harnesses will naturally optimize for their own model families. Any enterprise or sophisticated setup requiring multi-vendor pipelines—mixing and matching models based on task strength, cost, or redundancy—requires a neutral orchestrator. Most teams benefit from avoiding single-vendor lock-in at the workflow layer. Complex Topology & Orchestration: Native clients handle tight, local task execution well, but they aren't built for complex multi-runtime topologies. I outlined how I use OpenRig for persistent, higher-complexity setups here: #65 (comment). After running OpenRig alongside several other meta-harnessing tools, I put together my own comparative analysis. I’ve attached the analysis below for reference (note: this is independent, not official OpenRig material). harness-orchestration-layers.pdf |
Uh oh!
There was an error while loading. Please reload this page.
I've been really excited checking the project out and testing it out, liking it so far. Saw the youtube vid and it articulated and addressed a bunch of the issues I've hit trying to push LLMs to do more.
I'm wondering - My experience from the beginning of LLM's being able to write any code has been that anytime there's a shortcoming in the model's intelligence, I find myself trying to make scaffolding around it that corrects for that deficiency of intelligence. But then the models outgrow the need for it, and my prompts can get shorter and shorter, my scaffolding looser and looser, everything gets more natural.
OpenRig seems to me to be that sort of thing (scaffolding to correct for lack of model intelligence). Is that true? If so, If models were intelligent enough, would the queue and its structural guaranteeing of task handoff/completion, the prevention of bureaucracy, the proper scoping of context, etc, all be necessary?
If that's true, then for frontier models, the scaffolding rigor of OpenRig may be unnecessary in a year or two. (but would continue to be relevant to smaller/faster/cheaper/local models for a while after)
I'm just wondering if the fact that models will keep on getting smarter is part of the vision/awareness of this project. Not really sure if it matters, just curious.
Well I guess it could materialize later on as different 'modes' that OpenRig can run in, depending on model intelligence. Frontier: looser (fewer tokens to scaffolding overhead), small llm: full rigor (more token overhead, but enables small llms to be coherent on large and long time-scale codebases)
All reactions