Yesterday's Top Launches: 5 Tools from October 11, 2026
Five launches from October 11 mostly focus on keeping AI agent output from disappearing after a session, with additional tools for robotics physics and sales outreach.
Five launches from October 11: agents get memory, cheaper models, and better references
Yesterday’s crop of new developer tools shared an obvious obsession. Four of the five exist because someone got tired of watching AI output disappear the moment a session ends, and the fifth exists because robots are getting better at physics. There is also one sales tool in the mix, since outbound teams apparently want the same treatment. Here is what shipped, what each one actually solves, and where the rough edges are.
Busabase
The problem Busabase names is one most teams recognize but rarely fix: agent work scatters. A quote drafted in ChatGPT, a cleaned customer list in Claude, a script in Cursor before the author changed jobs, a meeting decision in Gemini. Each artifact lives and dies in a personal chat window. Busabase is an MIT-licensed database and workspace that gives those artifacts a home, organizing everything into four Bases for data, knowledge, apps and skills. A record carries its body, its sources and its review history, so an agent writing to a table produces something a person can open tomorrow, and vice versa.
What stands out is the review layer. Every write records who made it and what changed, and important changes can be held for human approval before they touch canonical data. It runs as a desktop app, via Docker, or in the cloud, and connects through MCP, OpenAPI, the CLI or Agent Skills, which means Claude Code, Codex and Cursor plug in without anyone switching tools. The free tier covers a single space with three seats; Plus runs $10 per seat monthly.
The honest caveat is scope. Calling yourself the general system of record for AI agents is a big claim, and the product’s success depends on teams actually abandoning chat as their default scratchpad. That is a habit change, not a feature.
Together Link
Together AI’s contribution is smaller in ambition and easier to justify. Together Link is a free, open-source bridge that points the agent harness you already use, whether that is Claude Code, Codex, OpenCode or Pi, at open models like Kimi K3, GLM 5.3 and DeepSeek V4.1 Flash. You keep your login, settings and history. Nothing gets rewritten.
The Auto Router decides which model handles a task by reading the first prompt of each session. Easy work routes to something cheap, harder problems go to a stronger model. Every session prints its token and dollar totals on exit, and a usage command shows the last seven days. The company claims 50-80% savings against running everything on Opus 5.5, which is plausible given the per-token spread.
It is in beta, macOS and Linux only, and the tool is free while the models are not. Worth trying if your coding agent bill has quietly become a line item.
Odyssey 3
Odyssey-3 is the outlier. It is a foundation world model, built as an autoregressive diffusion transformer, that generates embodied environments from a prompt and predicts how they change as an agent acts. The founders spent a decade on driverless cars, and that shows in the priorities. Physical accuracy is the point.
The numbers are strong on the team’s own benchmarks. Odyssey-3 Pro scored 66.1 on Physics-IQ Verified’s video-to-video test, the highest reported, and the model handles first-person and third-person navigation plus independent camera movement. More interesting are the adaptation results: with tens of hours of robot demonstrations it completed manipulation tasks and recovered from missed grasps, Flexion built humanoid control policies on it that beat tested VLA baselines, and a driving policy trained on 20 hours of Indian road data drove in closed loop with the backbone frozen.
These are self-reported figures, and the benchmark caveat is real. Access is a research preview you have to request.
AgentSDR
AgentSDR is an open-source, self-hosted AI SDR workspace that runs email, LinkedIn and WhatsApp from one place and pairs it with an AI CRM. It positions itself against Clay, Smartlead and HubSpot by putting reach, triage and measurement in a single system you host yourself.
The channel guardrails are properly thought through: per-mailbox daily caps, sending windows, randomized gaps between LinkedIn invites, warm-up periods for new WhatsApp numbers. Every reply gets classified, every draft waits for approval in an “Action required” queue, and the AI can only move a lead forward on its own, never back. Your OpenRouter key, your Postgres, your storage bucket.
What you are signing up for is sysadmin work. Docker Compose gets you running quickly, but Google Workspace domain-wide delegation and Unipile accounts are not a five-minute setup. For agencies running many clients in one deployment, the trade may be worth it.
Refs
Refs takes the most quotable premise of the day: Claude is only as good as the references you give it. The library holds 3,032 films, each broken down shot by shot with cut maps, cuts-per-minute figures and average shot length, then written up as a 15-plate blueprint and a runnable prompt recipe. An MCP server lets Claude Code, Codex or Cursor search the archive and pull blueprints into context.
Notably, it does not host the videos. It publishes only measurements and analysis, crediting every source, which keeps the method teachable without copying footage. Browsing is free; the full plates, recipes and MCP access sit behind Pro at $10 monthly billed yearly.
Five launches, one recurring theme: give the agent somewhere to put what it learned.