What happened
The GitHub release page says Tucano is a team of AI agents that runs on a user's computer and works through channels. The agents can research, write, code and use company tools. Tucano runs on supported command-line tools—including Claude Code, Codex, Cursor, OpenCode, Pi, Antigravity, Grok or Cline—using the user's own subscription. Installation requires a Mac, Linux PC or Windows 10/11 PC, and at least Claude Code or Codex installed and signed in.
The headline addition is Goals, invoked with /goal … in a channel (the release also documents /objetivo …). Agents pass the turn to one another, for example to build, review, fix and review again, until one calls concluir_objetivo with a result. Each reply receives the goal and the number of replies remaining. The loop ends at 12 replies, if an agent repeats itself, fails to pass the turn or finish, or if a reply fails, is stopped or reaches its limit. The channel receives a line explaining why the goal stopped. Tucano stores a goal in memory, but the release says it ends if the app closes.
A separate “Keep going on its own” setting lets a response continue after its per-reply limit when it was making progress and a tool call succeeded. It can happen up to two times consecutively, provided the monthly limit can cover another reply of that size; otherwise the agent waits for a person to continue it. The setting is on by default. When the app reopens, replies interrupted by closing resume automatically unless the workspace is paused or more than 12 hours have passed.
The agent browser can download files into its own folder, up to 20 MB each, and attach them to the conversation. An administrator can save logins on an agent's page, stored in the system keychain or as a 1Password reference. The browser fills a saved login only at the exact saved address and over HTTPS, except HTTP on localhost; passwords are masked in snapshots and screenshots and never shown to the model. For verification codes, the agent reads connected email or asks the user.
The release also changes how actions are applied. Changes Tucano proposes inside the app—such as agents, teams, channels, routines and agent settings—take effect when its reply finishes and can be undone for 24 hours. Connector access, the monthly limit and changes the user could not apply themselves still wait for confirmation. Agents without a team or lead can work as colleagues, and multiple agents called in one message respond in sequence, each reading what previous agents said.
What the archive adds
New Runtime covered Tucano 0.2.2's experimental self-hosted server mode and Gemini CLI support on 2026-10-31. That was an earlier release of the same product; 0.2.41 is a new event, adding goal-scoped work and browser and workflow controls rather than restating the server and harness changes.
Trend signal
This release supports the shift toward bounded agent loops: major agent platforms shipping primitives for goals, budgets, schedules, subagents and stop conditions. Tucano's 12-reply ceiling and explicit stop cases make the goal more than an open-ended chat, though the app-closing behavior means the loop is not yet durable across that interruption.
Why now
The release brings goal execution and browser access into the same local, channel-based agent team. That combination matters when teams want agents to act across tools without letting a long-running task proceed without limits: this version pairs handoffs with stop conditions, budget-aware continuation and approval gates for connector access.
Why it matters
A goal loop shifts the operator's job from manually prompting each agent turn to defining the outcome and watching the exceptions. The reply ceiling and stop conditions provide a basic boundary for unattended work; the automatic continuation rule adds a narrow recovery path for a response that was already making progress. Neither removes the need to inspect results, but both make delegation more explicit than a sequence of unbounded chat turns.
Browser access raises a separate operational question: how to give an agent access to authenticated sites without exposing credentials to its model. Tucano says saved logins are constrained to an exact address over HTTPS and passwords are masked. The confirmation boundary also distinguishes reversible in-app edits from connector access and monthly spending limits. Teams evaluating the release can therefore assess not only whether agents can complete a task, but where the system pauses for human authority.
Limits and uncertainty
The release describes product behavior, not measured task-completion rates, failure rates or the quality of agent-produced work. It does not establish how reliably agents recognize repetition, choose to pass a turn or decide that a goal is complete. Goals end when the app closes, so this is not a durable background job across that event. Automatic continuation is capped at two retries and depends on the monthly budget. The release does not quantify browser-login security beyond its stated address and transport restrictions, and those controls have not been independently tested here.
Skill to build now
For a bounded trial, write one channel goal as a verifiable outcome rather than a vague instruction, then observe whether agents can build, review and correct it within the 12-reply ceiling. Test the stop behavior deliberately: include a case where a reply does not pass the turn, and check that the channel explains why work stopped. For authenticated browsing, start with a non-sensitive account and confirm the exact saved HTTPS address before expanding access.
Reader takeaway
Tucano 0.2.41 makes multi-agent work more autonomous, but its most useful design detail may be the boundaries around that autonomy: reply caps, budget checks, stop conditions, scoped saved logins and confirmation for connector access. Treat goal mode as a bounded delegation loop, not as a persistent job runner.