How the DeepSeek agent harness works
A chatbot answers. MicroLearnerAI is an agent harness: a DeepSeek harness that plans, writes code, runs it, checks it, and hands you the finished file. If you've searched for a "DeepSeek agent harness," a "chat harness," or a "code execution harness," this is the concrete version — what actually happens between your message and the artifact.
What "harness" means here
An agent harness is the software layer around a model that lets it take real actions — plan, call tools, execute code, read the result, and retry — instead of only producing text. Where DeepSeek's own chat is the model talking, MicroLearnerAI is the harness that turns the same model into a worker. The DeepSeek agent harness is what makes "build me a spreadsheet" end in a real .xlsx rather than a code block.
A code execution harness is the part most chats don't have: a sandboxed runner where generated code actually executes, server-side, with a review-and-approve step before anything runs. And because the same harness can be triggered on a schedule, it doubles as an agent loop harness — the engine behind agent loops that fire while you're offline. Same chat harness, three entry points: a message, a schedule, or an inbox.
The pipeline at a glance
Every message goes through the same loop:
- Understand & plan. The model turns your request into a concrete plan — what to build, what file type, what data.
- Pull context. Attachments are read (Office files into context, images described by a vision model), and relevant skills are fetched.
- Write code. The model writes the program that will produce your file — a Python script, for example, that builds an Excel workbook.
- Execute server-side. The code runs on the execution engine (sandboxed serverless compute), not your machine. You never need Python, a compiler, or a library installed.
- Verify. The run's output is checked — did it produce a file? Is it the right shape? — before it's presented as done.
- Deliver the artifact. The working file is attached to the message and saved to your Settings → Artifacts gallery.
Watch it happen: the status bar
While a run is in flight, the chat shows a live status bar with pipeline stages — memory/context → AI processing → response building. That's the harness working: context assembly, then model inference, then building and attaching the output.
What the harness can build
The execution engine produces real, downloadable files that your regular software can open:
- Excel workbooks (
.xlsx) with formulas, formatting, and charts - PowerPoint decks (
.pptx) - PDFs
- PNG images & charts
- CSVs
- HTML pages / small tools that run in a browser
Files appear as artifact cards in the chat with syntax-highlighted code blocks. Edits show in a diff viewer (red/green, intra-line) with per-file version history.
The models behind it
All reasoning runs on DeepSeek (api.deepseek.com), currently the deepseek-v4-flash family:
- Instant — default, fastest, best for everyday builds
- Expert — deeper reasoning for harder problems
- DeepSeek v4 Pro — available on the Max plan and above
Images you attach are understood via a vision model, and Office attachments are read server-side into context.
Why usage is cost-based, not per-message
Because a "build me a spreadsheet" request can burn far more tokens than a "hi" — the model writes, runs, and iterates on code — usage is metered in model cost, not message count. Each plan has a session cost cap (USD of API cost) that resets every few hours; within the cap, messages are unlimited. See Plans & pricing.
The same harness, unattended
The identical pipeline powers scheduled tasks (recurring jobs that run the loop without you) and the mail bridge (email a task, get it done while the app is closed). Same planning, execution, and artifact delivery — just triggered by a schedule or an inbox instead of a chat message.