{"slug": "idkmesh-what-if-ai-agents-had-to-prove-their-work", "title": "IDKMesh: What if AI agents had to prove their work?", "summary": "A developer is building IDKMesh, an open-source research and engineering project that separates AI agent work generation from independent verification, treating worker output as candidate results rather than accepted completions. The project introduces Work Units with explicit boundaries and security constraints, replaceable workers, and verifiers that produce their own evidence artifacts, with the goal of preventing verification bottlenecks as agent swarms scale. A tool called `idkmesh gate-audit` is already available to measure how many independent votes a review panel is actually worth, while the full Verified Swarm Runner remains a research foundation rather than a production-ready platform.", "body_md": "[A few days ago I wrote about one result from IDKMesh: a panel of 25 verifiers that, after measuring their shared errors, was effectively worth only one independent verifier.](https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Frr7mgdxeketjk1hdfupa.png)\n\nThat experiment is interesting, but it starts in the middle of the story.\n\nSo this post is the simpler version:\n\n**What is IDKMesh, why am I building it, and how is it supposed to work?**\n\nAI agents are getting very good at producing things.\n\nThey can write code, edit files, call tools, generate tests, open pull requests, and sometimes operate infrastructure.\n\nBut there is a basic problem:\n\n**An agent saying “I finished the task” is not the same as evidence that the task was done correctly.**\n\nIf we scale from one agent to ten agents, or one hundred agents, generation becomes faster.\n\nVerification can become the bottleneck.\n\nAnd if all of those agents or reviewers make similar mistakes, adding more of them may only give us more confidence, not more truth.\n\nThat is the problem IDKMesh is exploring.\n\nIDKMesh is an open-source research and engineering project about coordinating:\n\naround uncertain work.\n\nThe name comes from a simple idea:\n\n**I don’t know. You don’t know. Together, the mesh can discover, build, verify, and learn.**\n\nFor me, the most important word is not “mesh.”\n\nIt is **verify**.\n\nIDKMesh tries to separate the thing that does the work from the thing that decides whether the work should be trusted.\n\nImagine I give an AI agent this task:\n\nFix a bug in a Python project.\n\nA normal agent workflow might look like this:\n\n```\nTask\n  ↓\nAI agent\n  ↓\nCode change\n  ↓\n“Done”\n```\n\nThat is useful, but the same system that created the answer is also telling us the answer is good.\n\nIDKMesh is aiming for something more like this:\n\n```\nHuman goal\n    ↓\nBounded Work Unit\n    ↓\nOne or more replaceable workers\n    ↓\nCandidate results\n    ↓\nIndependent verification\n    ↓\nRecorded evidence\n    ↓\nHuman / governance decision\n```\n\nThe worker can propose.\n\nThe verifier can check.\n\nThe evidence can be inspected later.\n\nAnd the final authority stays separate.\n\n**Worker success != acceptance.**\n\nThat small rule changes a lot.\n\nAt a high level, IDKMesh uses a few simple building blocks.\n\nInstead of giving an agent unlimited authority over a project, work is described as a **Work Unit**.\n\nThe task says what needs to be done, what the boundaries are, and what security constraints apply.\n\nA worker is allowed to try.\n\nThere can be multiple attempts.\n\nThe first answer does not need to become the final answer.\n\nWorkers are replaceable. Their output is a candidate.\n\nA verifier should not simply trust the worker’s own completion claim.\n\nVerification has its own plan and produces its own result.\n\nThis is where tests, independent reviewers, policy checks, or other evidence can enter.\n\nIDKMesh treats provenance and verification results as real artifacts, not temporary chat messages.\n\nThe goal is to be able to ask:\n\nA worker does not get merge authority just because it says “success.”\n\nA verifier recommendation is also not automatically the final decision.\n\nThe integration step stays explicit.\n\n**[Place the IDKMesh explainer image here.]**\n\nThis part is important because I do not want to describe research as a finished product.\n\nThe full **Verified Swarm Runner is not yet a production-ready multi-agent platform**.\n\nThe repository is still a research and engineering foundation.\n\nBut one useful tool is already available: **`idkmesh gate-audit`**.\n\nIt takes verdicts from a review panel you already have and asks a practical question:\n\n**How many independent votes is this panel actually worth?**\n\nThe repository also contains experiments, schemas, validators, simulations, interoperability work, and pieces of the future runner.\n\nIf you have Python 3.11+:\n\n```\ngit clone https://github.com/MSKazemi/idkmesh\ncd idkmesh\n\npython -m venv .venv\nsource .venv/bin/activate\n\npip install .\nidkmesh gate-audit examples/gate-audit/panel-votes.example.json --pretty\n```\n\nYou can also run the repository demo:\n\n```\npython scripts/demo.py\n```\n\nNo model API key is required for that demo.\n\nThe long-term idea is bigger than a review tool.\n\nI want to explore whether large groups of humans and AI agents can work on real problems without making trust an afterthought.\n\nThat means learning how to:\n\nI do not know yet how far this model can scale.\n\nThat uncertainty is part of the project.\n\nThe repository is designed to test those questions rather than pretend they are already solved.\n\nBecause a system about verification should itself be inspectable.\n\nIf IDKMesh makes a claim, I want the experiment, code, result, or limitation to be visible.\n\nAnd I especially want other people to find the assumptions that are wrong.\n\nIf you are interested in AI agents, verification, distributed systems, CI, developer tooling, or human-in-the-loop systems, take a look at **MSKazemi/idkmesh** on GitHub.\n\nYou do not need to understand the whole project to contribute.\n\nEven showing where the newcomer path is confusing is useful evidence.\n\n*This article was prepared with AI assistance and manually reviewed against the current IDKMesh repository before publication.*\n\n**Suggested DEV tags:** `#ai` `#opensource` `#python` `#agents`", "url": "https://wpnews.pro/news/idkmesh-what-if-ai-agents-had-to-prove-their-work", "canonical_source": "https://dev.to/mskazemi/idkmesh-what-if-ai-agents-had-to-prove-their-work-5bg4", "published_at": "2026-09-21 19:15:11+00:00", "updated_at": "2026-09-21 19:25:12.160487+00:00", "lang": "en", "topics": ["ai-agents", "ai-safety", "ai-research", "developer-tools", "ai-tools"], "entities": ["IDKMesh", "Verified Swarm Runner", "idkmesh gate-audit", "GitHub", "Python"], "alternates": {"html": "https://wpnews.pro/news/idkmesh-what-if-ai-agents-had-to-prove-their-work", "markdown": "https://wpnews.pro/news/idkmesh-what-if-ai-agents-had-to-prove-their-work.md", "text": "https://wpnews.pro/news/idkmesh-what-if-ai-agents-had-to-prove-their-work.txt", "jsonld": "https://wpnews.pro/news/idkmesh-what-if-ai-agents-had-to-prove-their-work.jsonld"}}