Agentbits

Release notes

What changed in each version of Agentbits, in plain words. The newest is at the top.

v0.3.0Latest

Know what needs you, and trust the result

A single place for everything waiting for you, acceptance criteria the lead checks one by one, a "What changed?" view for each task, and fixes for six problems an independent review found.

New

  • Needs you (top bar): approvals, work waiting to be merged, reviews the lead couldn't decide, tasks on hold with the reason, agents stopped by an error — one list, one click to the task.
  • Cards say why a task doesn't move: "Waits for TA-1 to be merged", "On hold: failed twice …", "Your decision", "Waits for race partner TA-7".
  • Acceptance criteria for a task ("Tests pass", "UI text in Turkish", "Don't touch db/"). The assignee sees them; the lead checks each one and sends the task back if one isn't met. The lead can add them when it plans tasks.
  • What changed? in a task (projects with "Each task in its own git working copy"): changed files, added and removed lines, and the diff.

Fixed

  • An agent's work could be lost when git couldn't commit it (for example because of a signing setting). The task now waits with the reason, and a working copy with unsaved changes is never removed.
  • When a task's own git working copy couldn't be prepared, the task ran in the shared project folder. It now stops and tells you why.
  • A task could start before the code it builds on was merged into the project folder. It now waits for the merge.
  • If the first 25 tasks in the queue were on hold, the rest never started.
  • An edited, replaced or deleted decision stayed in memory in its old form; memory now follows the decision.
  • With OpenAI-compatible and Anthropic providers, an error in the middle of an answer or a dropped connection could count as finished work. Both are errors now.

To upgrade, run the new installer over the old one; your data is kept.

v0.2.0

Teams that work in parallel

The lead talks to its team, duplicated developers share the work without overwriting each other, and you can share whole teams with friends. Plus Gemini and local models, old Office files and OCR, web reading, semantic memory and safety limits.

Teamwork

  • The lead talks to the team: it hands out work, sees what members finished and leaves notes on their tasks.
  • Parallel developers: duplicate an agent into twins; the lead splits work between them, tasks can wait for each other, and everyone sees who works on which files.
  • Each task in its own git working copy (project setting): tasks work on their own branch and are merged when approved.
  • Race mode (optional): a task also goes to a member on another provider or model; the lead keeps the better attempt.
  • Share teams and agents as .agentbits files; import from a file or a GitHub link with a full preview.

Providers, tools, files

  • Google Gemini, Ollama and LM Studio through the OpenAI provider's compatible endpoint.
  • Codex CLI agents get connection, file and app tools.
  • Agents with the web permission read public pages as text.
  • Custom reports, skills from private GitHub repositories, MCP resources and prompts.
  • Old Office files (.doc, .xls, .ppt) and .rtf; offline OCR for scanned PDFs and images (English, Turkish).
  • Optional semantic memory: notes are also found by meaning.

Safety and polish

  • Per-run time, token and cost limits; a reviewed list of environment variables passed to CLI agents; secrets masked in transcripts.
  • Turkish UI: tab names and Kanban columns are Turkish too.
  • Demo mode simulates research, writing, data and code tasks differently.
v0.1.0

First public release

Build your own team of AI agents — a lead plus specialists — hand them work and watch them do it in a pixel-art office.

In this release

  • Your own AI organization: teams, roles, per-agent permissions and six ready-made templates.
  • A project lead that turns your request into tasks, assigns them and reviews the results.
  • Live pixel-art office, Kanban board and a terminal for every agent, with token usage.
  • Project memory and a decision log.
  • Connections to SSH servers and PostgreSQL, MySQL/MariaDB and SQL Server: denied by default, Production changes need your approval, every call is audited.
  • MCP apps and installable skills; English and Turkish UI.
  • Claude Code CLI, Codex CLI, Anthropic API, OpenAI API and a free Demo mode.