What a development team looks like today
A development team is still made of people. What changed is what sits next to each of them: an agent, and around the agent a small workplace the developer built for themselves.
Key numbers
- 2×
- as many code changes per engineer, while company metrics stay flat (Faros AI, 10,000 developers)
- 59%
- of their work is what Anthropic’s engineers say they do with Claude (Anthropic, December 2025)
- 0–20%
- is all that more than half of them can fully delegate to an agent
In this article
Many developers have an agent now
Look at a team today and the shape is often the same: a shared repository, a shared tracker, and next to many developers an AI agent. Faros AI analyzed the work of 10,000 developers in 1,255 enterprise engineering teams. It found engineers completing twice as many code changes, while company metrics stayed flat (Faros AI, 2025). Anthropic studied its own engineers (Anthropic, December 2025). They say they use Claude in 59% of their work, yet more than half said they can fully delegate only 0 to 20% of it.
Two things follow. Code comes faster. But a person still has to check what an agent did, so the limit looks like the developer’s attention, not the agent.
The workplace each developer builds
So developers set up a corner of their own. A notes file the agent reads at the start of a session. A plain-text task list in the repository, so the agent knows what is next. A script that starts the same three sessions every morning. What I keep hearing from developers: people even pass tasks to each other through files, next to the official tracker.
None of this replaces Jira or whatever the company uses. The tracker is the company’s record of who owes what. The personal layer is something else: how I keep my agents busy and know what they did.
- Dev Aclaudemy corner
- notes.md
- todo.txt
- start.sh
- Dev Bcodexmy corner
- CLAUDE.md
- tasks.txt
- Dev Copencodemy corner
- plan.md
- run.sh
- notes.txt
That layer, done properly
harnsy is that personal layer, built so you do not have to improvise it. It runs on your machine and connects Claude Code, Codex and OpenCode. Agents from different tools can message each other and work as one team. Each has a role, a lead hands out the work, and every project has a task board. When an agent’s context (everything it holds in one session) is full, the work moves to a fresh agent with written handover notes. When an agent needs your answer or your permission, the request shows up in one place.
You2 wait for you
- Leadclaudeworking
- Engineering
- Backendcodexworking
- Frontendclaudeworking
- Quality
- Revieweropencodeidle
Not connected: tracker connectors are on the roadmap.
- 1You“Add a CSV export to the reports page”→ Lead
- 2LeadAll columns or the visible ones?AllVisiblewaits for you2You“Visible”
- 3LeadTask #452 with criteria and a checklistNew→Taken
- 4BackendIn progressown git worktree (git practice)4FrontendIn progressown git worktree (git practice)
- handover
When an agent’s context fills up, it writes a handover note and a fresh agent takes the role.
- 5Frontendready for review→ Reviewer5Reviewerchecked: 4 of 4
- 6YouIn reviewAcceptReturnwaits for you
- 7YouAccept →Done
Your tickets stay where they are. harnsy does not read or write Jira: it keeps the working layer next to it. Tracker connectors are on the roadmap, not in the product.
More than one machine, still one owner
Sooner or later one machine is not enough: a laptop you carry, a server where long runs live, a desktop at home. Since 0.9.0, with harnsy Max they join into one cluster. One machine is the main node; the others are members. The main node sees its members’ agents. If a member allows it, the main node also sees their terminals and can type into them. If the main node has no public address, the members reach it through our hosted relay (beta). The links between machines use TLS by default. The relay only joins encrypted streams; it sees which node (as a hash), the network address it connects from, when and how much, not what is inside.
A practical pattern: give each agent its own git worktree, so two agents do not touch the same files. That is a git practice, not a harnsy feature, and it works the same on a server machine as on your laptop. The repository can live on the server; you watch the agents from the laptop.
One owner: youharnsy Max
TLS · the relay joins encrypted streams
Where harnsy stops
harnsy has one person in charge: whoever owns the machines. It does not give your colleagues accounts, and it does not merge work between people; that stays in git, your review process and your tracker. The cluster is one person’s several machines, not a shared workspace for a company.
What you get is the part that is yours: your agents, your roles, your task board, your handover notes, on your own machines.
Start with one project
Paste one line into your agent to install harnsy: harnsy.dev/install. Without a key: 1 project, 10 agents at once, no other machines. Several machines are harnsy Max.
Start with one project
Your agents, your roles, your task board and your handover notes, on your own machines.
Install harnsyharnsy is not affiliated with Anthropic, OpenAI or OpenCode; their names are theirs.