deepgent
A domain-locked autonomous engineering agent for autonomous vehicles, computer vision and embedded systems. It takes a task from natural language to a verified artifact running on target hardware at spec.
"Tests pass" is not done
General coding agents optimise for a green test suite. For edge autonomy work that is the wrong finish line: the artifact has to run on the target board, inside a power and thermal envelope, at a stated frame rate and tail latency, without losing accuracy to the quantization that got it there.
Definition of done
The artifact runs on target hardware and meets the stated metric: fps, p99 latency, mAP delta, memory, thermal. "Compiles and tests pass" is an intermediate state, never completion.
deepgent is also deliberately domain-locked: AV, CV, embedded and robotics-adjacent edge AI only. Narrow scope is what makes the hardware-aware reasoning (board compatibility, errata, CUDA versions, BOM tradeoffs) possible at all.
From a sentence to a board
Built on the Claude Agent SDK. The capability surface is exposed both as a CLI and over MCP, so the same tools drive the agent loop and any MCP client.
What the agent can actually do
| Area | Capabilities |
|---|---|
| Hardware | Host profiling and doctor checks, hardware conflict detection, CUDA compatibility checks, board catalogue, errata scanning, BOM advice |
| Models | Model selection, quantization sweeps, accuracy scoring and accuracy gates, differential comparison |
| Profiling | Latency profiling, Nsight profiling, thermal profiling, soak tests |
| Generation | ROS 2 node generation, systemd unit generation, driver scaffolding |
| Validation | Shadow mode, replay, bisect, premortem, reflection |
| Fleet | Fleet state, triage, telemetry summaries, upgrade checks |
| Skills | Skill authoring and evaluation, with skill-lift verdicts |
Twelve surfaces on one event loop
A dense, dark, Blender-style desktop app over the full capability surface, shipped as an optional extra so the core wheel stays dependency-light:
uv pip install 'deepgent[gui]' then deepgent gui
It runs the whole UI on a single qasync event loop, so deepgent's async core (orchestrator, board runners, soak, evals) executes without freezing the UI and without worker threads. Twelve surfaces cover running tasks, profiling, model and performance analysis, compatibility reasoning, and knowledge.
