The problem
Almost right is still wrong. #
One AI model sounds just as sure when it is wrong. You find out later, usually after you have used the answer.
Agent Forge does not ask you to trust one model. It puts a team on the job, and a different model checks what the team found.
66% of developers name it their top AI frustration. Stack Overflow survey, 2025
One ask. Ten agents from four AI labs. One catch. #
A real turn from 16 Oct 2026, replayed from its own record.
The same turn, on the platform’s own screens
This happened in production. #
On 16 Oct 2026 we asked it to re-check an open-source database’s licence, benchmarks and hosting cost from that day’s sources. Ten agents from four AI labs took the job.
One worker reported that it could not read the latest release. A checker from another lab found that claim contradicted by the team’s own stored result, and the final answer said so.
Real work
What you can hand it. #
The same team runs every workspace. In each one it knows where it is and what it can touch.
-
Research you can stand behindAsk for a vendor comparison, a due-diligence note or a market check. Workers fetch current sources, and a different model tests the claims against what they found. How claims get checked →
-
Answers from your dataConnect Postgres, MySQL, SQLite or MongoDB, or upload a spreadsheet, and ask in plain English. It writes the query and shows you the result. Any change it wants to make waits for your approval. Data Grid and the other workspaces →
-
Commercial trucking underwritingUpload a submission. It pulls out the fields, checks the carrier’s federal safety record, and answers from a research library. Each rating factor names where it came from, and the quote stays a recommendation for the underwriter. See it in the tour →
-
Images and videoMake an image, edit one, or turn it into video. Direct Media Studio in chat or set the controls yourself, and a simple image stays a quick job. How each part gets its model →
-
SoftwarePlan a change, write it, and read what will change before any file moves. Connect GitHub, Notion or Linear and the work reaches your own tools, with your approval. See the approval step →
-
Reports, decks and spreadsheetsAsk for the finished file: Word, PowerPoint, Excel, PDF or a web page. It lands in your workspace ready to download. Start with the docs →
Under the hood
What happens when you hit send. #
Not every request needs every step. A quick question takes the short path.
-
It reads the jobWhat kind of work this is, how hard it is, and what it needs. The pipeline →
-
It remembersRelated chats, files and past results come back by meaning, not by keyword. Memory →
-
Leaders planHard, new work goes to a council of up to three top models. Simple asks, and work it has solved before, skip it. The council →
-
Agents split itEach part goes to the model best placed to do it. Agents can spawn more agents. How agents spawn →
-
Another model checksWhen a turn used tools or agents, a different model tests the result against the stored evidence. What fails is corrected, or marked unverified. How the check works →
-
It learnsThe outcome is scored, and that record shapes who gets the next similar job. How models earn trust →
You stay in charge
Nothing changes without you. #
Agents read freely. Anything that changes your files, your data or your connected apps waits for your yes.
- You approve the exact action. The card names what will change, and you get the diff back.
- You see every call. Under each reply: the model, provider, tokens and cost of every call.
- It says what it could not confirm. An answer that fails its check is marked unverified, not passed off as done.
[See the approval step →](workflows_factual.html#approval)
[Where your data goes →](security.html)
It learns
It learns who to ask. #
Most tools pick a model by what it claims it can do. Agent Forge weighs what each model has actually done here.
Finished work is scored, and each model earns trust by what it delivered. Public leaderboards give a new model its start. Its record here does the rest.
How models earn trust → Built by
Prashant Dimri #
See the evaluation work behind Agent Forge: the tasks, the rubrics and the model tests.
Read Founder’s Corner → Live in private preview
The one-model era is over. #
Accounts are by request for now. Tell us what you are trying to build and we will show you the part that matters to you, on the real system.
Usually answered within a few days. Ask anything, including the parts that do not work yet.