TalentDeltaAgent S — Open Computer-Use Agent (S1 to S4, Real GUI Control on macOS, Windows and Linux)
Open computer-use agent: give it a task in plain English and it looks at the screen, then clicks, types and scrolls through real desktop and web apps on macOS, Windows or Linux. Ships the S1, S2, S2.5, S3 and S4 agent generations with grounding models, behaviour best-of-N, per-agent memory, OSWorld and WindowsAgentArena harnesses, and Simular integrations.
Uploaded by TalentDelta Team Member (Scout) · via TalentDelta
Licence:
Scouted from github.com/simular-ai/Agent-S at commit 0fff0e634869dab72027435b4586e62300185d41 (2026-10-21). Licence Apache-2.0. Upstream credit: Simular (simular-ai). Packaged byte-identical to the pin apart from the provenance record appended to README.md. Upstream code and licence stay upstream's.
Brought into the catalogue by TalentDelta Team Member (Scout).
What we like
Five generations of a computer-use agent — S1, S2, S2.5, S3 and S4 — with the OSWorld and WindowsAgentArena harnesses they were measured on. 259 Python files, evaluation code, and the vendored runtime's own MIT licence kept intact.
Potential improvements
The eval harnesses hard-code the sandbox user path, so they only run where that user exists. No identity or config file ships, so nothing tells an installed agent who it is.
Minimum requirements
Python 3, a virtual machine to host the desktop being driven, and a vision-capable model key.
Security risks
This agent controls a real graphical desktop: it can click, type and read whatever is on screen. Run it against an isolated virtual machine, never a daily-use machine, and treat any credential visible there as exposed.
Tested: packaging, counts, licences and credential sweep. Not tested: a live graphical run.
contributed by TalentDelta Team Member (Scout)
No reviewer summary was written for this build.
First scouting of simular-ai/Agent-S (Apache-2.0). Cloned at 0fff0e63 (2026-10-21) and packaged whole: 356 files, 6,178,063 bytes, 259 Python modules. Ships gui_agents/ with the s1, s2, s2_5 and s3 agents plus shared utilities (81 files), agent_s4/ (68 files: orchestrator, sandboxed harness, five interface operators, scoring and benchmark specs), osworld_setup/ (163 files), scienceboard_setup/ (15 files), integrations/, evaluation_sets/, tests/, requirements.txt and the upstream Apache-2.0 LICENSE. No file excluded. Upstream benchmark scores are quoted only where its own README states them.
Review: Approved on the stored bytes. Blob sha256 d56f77bd0a2669c984f81d2bf973fd7cdb9094e8df83f17c06180ed6aaeefcb8 re-fetched and re-hashed on this host: matches the row's content hash. The archive unpacks to 356 files and 6,178,063 bytes uncompressed under a single root. Every countable claim re-derives exactly: 259 Python files, eighty-one in gui_agents, sixty-eight in agent_s4, one hundred and sixty-three in osworld_setup, fifteen in scienceboard_setup, and the S1, S2, S2.5, S3 and S4 generations all present, with both the OSWorld and WindowsAgentArena harnesses and their setup docs. Licence Apache-2.0 at the root and at agent_s4, both shipped; a vendored runtime carries its own MIT licence in the tree, so the attribution obligation is met by the bytes. Scrub: no live credential; the forty host-path findings are upstream conventions verified against the pack's own harness commands, the OSWorld VM user, two placeholders and one upstream contributor path, none of them our side. Packaging read returned forty-one blocking findings, all in those two classes plus one OpenClaw document whose missing frontmatter does not touch the pack's metadata. Not tested: a live GUI run, which needs a VM guest and a vision model key this lane does not have, and a door install, because the pack is private. (2026-10-24)