One Agent Is Not Enough: Orchestrators, Subagents and Memory From Show HN
The first generation of coding agent tooling assumed one agent, one repository, one conversation. That assumption is now the bottleneck. Real work wants a planner and several workers, a reviewer that never edits, and memory that outlives the session. A batch of Show HN posts from the past month is entirely about that shift: orchestration, isolation, shared memory, and getting agents into the browser where the work actually happens.
These are the projects that treat your agent setup as a system to be arranged rather than a chat window to be prompted.
Orchestrating more than one agent
Running two agents against the same tree is a good way to lose an afternoon to merge conflicts. Pi Cowboy is a Herdr-native orchestrator for Pi that runs Codex, Claude Code, OpenCode and others as subagents, each inside an isolated copy-on-write Git worktree. The copy-on-write detail is the important one: isolation is what makes parallelism survivable, because a failed attempt costs you nothing to discard.
The other half of orchestration is credentials. Every vendor wants your agent to own the login, which means every integration is a new place to store a secret. Pi Codex Connectors takes the opposite view: it starts codex app-server on first use and lets Codex keep ownership of your login and connector credentials, exposing just three tools to list, inspect and call connectors instead of loading hundreds of schemas into context. Fewer tools is not just tidier, it is cheaper per turn.
Getting agents into the browser you already use
Headless browsers are fine until the task needs a session. Agent Chrome Relay lets your coding agents use your everyday Chrome with your real logins, driving background tabs rather than a fresh profile that has to authenticate from scratch. It is genuinely convenient and genuinely risky: an agent with your logged-in browser has your logged-in browser, so this belongs on tasks where you would have done the clicking yourself anyway.
The reverse problem is feedback. Telling an agent that the button is wrong wastes a turn on geometry. Spotnote is a Vite plugin that lets you annotate the running dev server and hand the note straight to the agent, which removes the most lossy translation in front-end work: describing what you can see.
Memory that survives the session
Agents start every session by re-deriving context they already had last week. IHMT Memory gives long-term memory to Claude Code, Codex and opencode as a tree of plain files. Plain files are the point: you can read them, diff them and delete the parts that are wrong, which is not true of an embedding store you cannot inspect.
At the far end of the same idea is the attempt to industrialise the whole path from ticket to merge. Engrams is an open source alternative to Devin, billed as an AI software factory for the enterprise and written in Rust. It is the heaviest item here and the only one that presumes an org rather than a person, which makes it worth watching even if you never deploy it.
Investigating rather than writing
Not every agent task is authoring code. Rea points agents at reverse engineering, from app behavior down to native binaries, which is a good example of agents being aimed at the tedious first pass rather than the final judgment: let the machine form hypotheses from observable behavior, then have a human confirm them.
Some projects push further and ask what the operating system would look like if the agent were the user. That question is easy to dismiss as a thought experiment until you notice how much of a modern desktop is arranged around one foreground window and a clipboard, neither of which an agent can use well. Veda is an agentic-native operating system written in Rust. It is a hobby OS and not a daily driver, but it is a useful thought experiment made runnable.
Two more worth knowing about
Serving is still the part that decides whether any of this is affordable at scale. DeepInfer is DeepSeek's open-sourced high-performance inference framework for its own models, which matters less as a curiosity than as a signal: the labs are now publishing the serving path, not just the weights.
And agents are leaking into personal workflows too. What makes it representative rather than cute is the constraint it accepts: everything runs locally, your data stays yours, and the agent does the part you were never going to do by hand. HN Hiring Ranker ranks every Hacker News Who is hiring post against your resume, drafts emails and prefills applications from a local dashboard running on Claude Code. It is a small tool that shows the shape of the bigger trend: the agent runs locally, owns your data, and does the part of the job you were never going to do manually anyway.
