Blog
Notes from running a team of AI agents: real numbers, plain words.
Competitor monitoring for small business: a weekly report on new products and prices
Name the competitor pages that matter. Every Monday a short report shows what is new on them and what changed in price, with links. You read it and decide.
Claude Code in Telegram: your AI lead’s questions, approvals and the weekly result in one chat
Your AI team stops when it needs your answer. Put your AI lead’s questions and the weekly result in a Telegram chat: press a button on your phone, and the work goes on.
Social media content calendar: a week of post drafts every Monday
Decide once what goes out on which day. Every Monday a dated calendar of post drafts for LinkedIn, X and Instagram waits for your review. You post them.
AI marketing agent with Claude Code: a small marketing team that drafts your week
Every Monday a plan for your week is ready to read: three post drafts and one newsletter or page idea, written by a small AI team. Free harnsy, one schedule.
Your AI agents get a browser: they open sites, you watch and take over
harnsy 0.11 gives your agents a browser on your own machine. They open sites, click and fill in forms; you watch the screen, sign in yourself and take control of any tab.
Claude Duet and a team of agents on your own machine
What Claude Duet is, as Anthropic and the launch coverage describe it, and where a team of coding agents on your own computer fits next to it.
Backups in harnsy: a copy before every update, and more if you turn it on
harnsy 0.11 makes a copy of its data before every update. See how scheduled backups, restore and the move to a new computer work.
Find any file your AI team made: the new Documents view
harnsy 0.11 groups every file your agents attached by task, kind or date, with thumbnails and search. See how the new Documents view works.
Inbox: client mail and chats arrive in your project, and you decide what reaches your AI team
Since harnsy 0.11, a project gets an Inbox: mail, a Telegram group and webhooks in one place, with rules for what goes to your AI team.
Letters: your AI team sends e-mail, and the first one to every address waits for your yes
Since harnsy 0.11, agents send e-mail by templates you approve, through your own mail server. See how Letters work and what you confirm.
Updating harnsy: it tells you about a new version, you press «Update»
harnsy never installs a new version by itself. See how the «Update» button works today and what the Updates page of harnsy 0.11 adds.
How to build an agent loop in harnsy
A loop is work that comes back by itself and gets better each round. Here is how to build one in harnsy from a schedule, a work item, one question to you and a note: our weekly LinkedIn plan is the example.
Hire a ready AI employee, or send us yours
hr.harnsy.dev is a catalog of ready AI employees for your harnsy team. Read one, hire it with one line, or publish your own with a pull request.
Claude Code, Codex and OpenCode in one team: give each a role
Claude Code vs Codex vs OpenCode without a verdict: how to set roles across the three, and what differs when they work in one team.
Assign who answers what, and your agent asks that person
In harnsy v0.10.0 you can give a post in your project, such as lawyer, to a person linked in Telegram. The agent’s question goes to that person, and only they can answer.
Claude Code code review with a second agent: what a reviewer caught
A reviewer agent checks another agent’s commit in a clean copy: 66 verdicts in one working day, a race found, one miss, and how to set one up.
Imagine your agents had a channel to your people
A way to think about a team of AI agents and the people who decide for it, and what harnsy does today when an agent needs an answer.
One project per job: how we run a company as parallel projects in harnsy
We run harnsy as four projects, each with its own lead and team of AI agents: how work crosses between them, what the human sees, and what got worse.
Claude Code auto-compact or a handover note: when each fits
What Claude Code’s auto-compact keeps and drops, how to steer it with /compact, and when a handover note and a fresh agent work better.
What three windows of our agents’ sessions taught us about handover points and token cost
We changed our agents’ settings and measured three windows of our own Claude Code sessions. What moves context and token weight: the handover point, waiting, messages, subagents.
Task management for AI agents: board and worktrees
What a task needs when an AI agent does it: criteria with evidence, an owner role, an order and a git worktree, shown on one real bug from our own work.
After the reboot: bringing a whole team of coding agents back, not one pane
Reboot, crash or update: how to get Claude Code, Codex and OpenCode sessions back, why tmux-resurrect isn’t enough, and how to restore the team.
Three machines, three people, one cluster
One main machine, members that join it directly or through a relay: what crosses between machines, what stays put, and what the relay can see.
Claude Sonnet vs Opus for coding agents: choose a model and effort per role
Sonnet 5.5 or Opus 5.5, and which effort? A role-by-role table, the prices with cache, and what we measured versus what is still a pilot.
You’re the CEO: why a hierarchy of agents is context engineering
Why companies use hierarchy, why it fits AI agents, and what our own numbers show about splitting context by role. Honest limits included.
AI employees: what works is a workplace, not a persona
“AI employees” is really a question of who is accountable. What vendors and studies say, and how we run agent teams by roles and handover.
Where our AI agents’ tokens actually go: a week of 65,000 requests
65,000 Claude Code requests in a week: 65% of the API-priced weight is re-reading the conversation. What that means for your usage limits.
What a development team looks like today
A development team is still made of people, each with an agent and a workplace of their own. How that personal layer looks, and where harnsy fits.