Most weeks the loudest launches are models. This week the more useful ones were the parts underneath: where an agent deploys, how it finds its next task, how it knows what your database looks like, and how you clean up after it. Eleven Show HN launches map that layer out, and they are unusually consistent about one thing - none of them want to replace the agent you already use.
One binary you can drive from a script
Radian is a coding agent shipped as a single Rust binary. There is no runtime to boot first, and the pitch is that context gets spent on the lines a task actually needs rather than on scaffolding. The detail that matters for anyone wiring an agent into automation is the headless command: it prints every step as a JSON line and exits with a code that describes how the run ended. That is the difference between an agent you chat with and an agent you can put in a pipeline, because a script can read the stream and branch on the exit code. The terminal session shows the same run as thinking, shell commands, reads and patches, each with timings.
Planning on the expensive model, everything else on the cheap one
Mixdog is an open-source desktop coding agent built around a cost argument. A lead model plans, while cheaper models handle search, edits and review. The project reports 63% fewer tokens than Codex CLI on the same model at the same success rate, which is its own benchmark and worth treating as a claim rather than a fact. The mechanism behind it is more interesting than the number: auto reasoning effort gives each message and tool step the effort it needs, judged in about 0.1 seconds by a 150 MB model that runs locally on the CPU. It is on by default, and it is the kind of small local classifier that shows up again and again in this batch of launches.
Somewhere for agent-built apps to live
Agents are good at writing small internal tools and bad at finding them a home. Brod is the self-hosted cloud your agent deploys team tools to, over MCP. You start it with docker compose, read a one-time sign-in link from the log, add it as an MCP server in Claude Code or Cursor, and ask the agent to deploy something; it answers with an address. Every app shares one sign-in, so nobody builds another user table, and apps read who is signed in along with their role and groups. Widefleet takes the same idea one step further and is an open-source platform for agent-built apps and workflows: your coding agent creates and manages apps through a CLI, and each app gets its own database and private file storage that survives code deployments. It also handles background jobs and schedules, and lets you deploy, preview, inspect logs and roll back.
A work board that lives in the repository
Pullboard is a work board that lives in your git repo. Agents pick up the next task, another agent checks it, and nothing merges until that second verdict lands. Three ideas do the work: Items each carry a bar they are checked against, Spec holds the plan one row per requirement where only the rows you approve count, and Doctrine holds your house rules with something that enforces each one, from a git hook to review. Every claim, submit, reject and accept is recorded in order, so nothing happens off the record. LangGraph Harness applies the same shape to GitLab and is a self-hosted AI coding agent platform built with LangGraph: it reviews merge requests from webhook events, resolves issues autonomously and chats with full project context. A React admin UI shows live runs, queue state and per-project analytics, including reliability, guardrail health and comment acceptance rate.
Keeping track of sessions you started somewhere else
Once you run more than one agent, the problem is knowing what is running. Faplex is a status list for the Claude Code, Codex and OpenCode sessions you already have, on every machine you can ssh to. It adds no hooks, plugins or wrapper, and installs nothing on the remote machines: it reads what the harness daemons already report, and opening a row hands your terminal the harness own attach or resume command, so every feature works exactly as it does without it. Sessions started anywhere show up, not just ones started from Faplex. Ghx solves the smaller, sharper version of the same problem for the GitHub CLI. It is a caching proxy where singleflight coalescing means several agents requesting the same PR list trigger one API call, and only allowlisted read-only commands are cached while mutations pass through and invalidate related entries. Context-aware cache keys prevent cross-repo collisions, no tokens are stored, and a dashboard shows hit rates so you can tune the TTL.
Cleaning up, and carrying your rules along
Two launches handle what happens after the merge. Fresh is an agent skill that starts you clean once a PR has merged: it syncs your default branch and removes branches and worktrees whose changes already landed, including squash merges, which is the case naive cleanup scripts get wrong. It shows you the plan first, stops on uncommitted changes, and keeps branches with unmerged work. Shadowclone goes at the other side of the loop and keeps your engineering preferences identical in every coding agent. It maintains one set of skills and instructions for Claude Code, Codex, Cursor, Pi, Antigravity and OpenCode, and can learn new preferences from the sessions you allow. It can also work while you are away: ask for a clone and it runs with your skills, on your own subscription, in Docker on your own machine, and opens a draft pull request as you.
Telling the agent what the database looks like
The last piece is context. LLMSchema extracts database schemas from PostgreSQL, MySQL and SQLite into compact Markdown, on the argument that an agent without schema documentation has to reconstruct the data model from application code and migration histories, which is slower and burns more context. It writes tables, columns, types, indexes, constraints and relationships into one portable document, with optional per-table files, and you can hand your coding agent the repository URL and ask it to install and configure it for the project. It is meant for development databases rather than production-critical documentation.
What ties them together
None of these launches try to be the agent. They are the deployment target, the board, the session list, the cleanup, the schema - the parts that decide whether an agent can do a day of real work inside a company instead of a demo in a chat window. The ones likely to matter most in six months are the ones that refuse to own anything: Faplex holds nothing your sessions depend on, Brod keeps no user table of its own, and Shadowclone is explicitly not a separate agent. That is the right instinct for infrastructure.