Writing code was never the hard part
Every AI agent can write code now. That was never the hard part.
The hard part is everything around it. Deciding what to build, and in which order. Working through the review. Fixing the red build. Reading the production errors nobody reads. And doing all of it without handing your laptop to a machine.
That is the part we built, and we built it inside a sandbox.
What to build, and in which order
An agent argues the change with you, against your architecture decisions, your domain model and the code itself, then drafts the epic, its tickets and the decisions worth recording. Nothing is published until you say so. Then the agents work on several tickets at once, each starting only when the tickets blocking it are merged. How plans work.
The reviewer, and the red build
When a reviewer comments, the agent works through the comments within minutes, and tells you what it fixed and what it disagrees with, with a reply ready to paste. When a ticket leaves a question open, it asks on the ticket and waits for whoever answers there. When the build goes red, one click starts the fix. Every step from plan to merge.
The errors nobody reads
An agent reads the new errors in your logs, checks each one against the code, and opens a bug with its root cause and a proposed fix. It becomes work only when someone decides it should. How triage works.
Without handing over your laptop
Agents that write code must run commands, and no list of allowed commands fits every repository. So every session gets a sandbox of its own, built for it and taken down when it ends, and around it layers that hold even when an agent is fooled.
- The sandbox. The agent cannot touch your SSH keys, your git configuration or your
.envfiles, and cannot reach the workspaces of your other agents. - The code host. The agent cannot reach it at all. The Runner pushes its branch for it, that branch and nothing else, never by force.
- The Runner. It connects out, and never takes a command from outside, only a description of the work.
- The tools. Each agent sees only the tools its job needs, and its access ends with its session.
If the sandbox cannot start, nothing runs. How the sandbox works.
Still your team's call
Your team decides what the agents work on, opens every pull request and merges. Tillerly runs on Claude Code, with your developers' own logins or an API key, so there is no new AI contract, and its agents reach your tools through MCP servers, with each developer's own access. Today it works with Jira, Bitbucket, Buildkite, Confluence and Datadog.