DeepSeek published its Super-agent prototype under the MIT licence. We installed their code unchanged, pointed it at a small repository with a real bug, and recorded 45 minutes of what it did. This page replays that run.
What is real: their code, every decision, ticket, commit, pull request and restart below. What is staged: the bug and the issues are ours, the Before screen is a simulation, and a rule named "bugs" was written by us. Limits we hit: it ran in observe mode, so proactive work was recorded and not carried out; two mentions were silently missed; the container kill was tested only with no uncommitted edits at stake. DeepSeek calls this a research prototype, not an official product, and says it differs from their internal version. Vizuara is not affiliated with DeepSeek.
You open a session for each task, paste in what it needs, and remember what each one knows. In DeepSeek's experience, a person can follow only 3 to 5 active sessions at once.
Nothing connects these sessions except you. A simulated screen, but you have probably lived it.
Press Play. On the left is what appeared on GitHub. On the right is what the prototype recorded, from its own database, reworded into plain sentences with the values unchanged.
DeepSeek Harness 0.2.1-rc.3 and the dsh-superagent-prototype code, in a Modal container, never on a personal machine, because its workers run with full access to the machine they are on. Workers used DeepSeek V4 Pro and the main session used V4.1 Flash, through OpenRouter. We switched off session-log upload and telemetry. Nothing was merged.
What did not work: mentions in the bodies of two issues were silently missed, probably a timing race in its GitHub polling, judging from its code. Comments and the receive-all switch worked. We also needed a workaround to sign its GitHub tool in. Its subagent feature was not exercised.
Built and written by Dr. Raj Dandekar and the team at Vizuara AI Labs, for a webinar on 4 November 2026. The design it follows is explained in full at superagent.vizuara.ai.
We teach with love, care and passion. If you want to build agents like this end to end, our live cohort Agentic Systems Engineering takes you from the LLM to retrieval, tools over MCP, the agent harness, workflows and evals, on a real business system. It is open to engineers moving into agentic AI roles and to enterprise teams.