Vincent v0.7.0 adds task notifications, CLI control and spending limits for unattended agents

Vincent, an open-source tool for orchestrating agentic development workflows, has shipped version 0.7.0, an update centred on operating agents without watching them. According to the release notes published on dev.to, the hard part of agentic tooling is no longer starting an agent — it is knowing what the agent is doing, whether it needs input, how to control it from a script or over SSH, and how to inspect it afterwards. The new version addresses those operational gaps directly.

A full-screen workspace for every task

The terminal interface received one of its largest usability overhauls so far. Opening a task now enters a dedicated full-screen workspace with separate views for steps and attempts, task details, output, and diffs. The selected attempt carries over between views, outputs can be switched across retries, and failed attempts show their result summaries directly in the timeline. The effect, as the notes describe it, is that the main task board goes back to being an overview while the detail lives one level deeper.

Agents can report their own progress

A running step can now publish a status message, for example vincent status "Running integration tests", which becomes visible in the TUI and is also exposed through the API. The author calls this a small feature that changes the experience of longer workflows considerably: instead of a generic running indicator, the status line can show what the agent is actually investigating or what it is waiting on. Agent steps also receive the VINCENT_* environment variables that command and check steps already had, letting agents report progress through their shell tools.

Notifications beyond the terminal bell

Previously, Vincent could ring the terminal bell when a task reached an interactive state — useful only if the TUI happened to be open. v0.7.0 adds configurable external notifications: a user-defined command runs whenever a task enters states such as blocked, awaiting_input, awaiting_gate or done, and receives a JSON payload describing the task and the state transition. Because the mechanism is command-based rather than tied to a specific notification provider, it can be wired into desktop notifications, Slack, webhooks, scripts, file drops or internal tooling.

Scriptable control without the TUI

A stated goal of the release was reducing dependence on the TUI for operational tasks. The vincent task command now covers actions including pause, resume, skip, approve, reject, retry, repair, archive and answer, which makes the tool practical from shell scripts, CI environments, cron jobs and remote SSH sessions. The notes include an example that lists blocked tasks as JSON and pipes their ids into a retry command one by one.

Debugging and remote inspection also moved to the CLI. vincent daemon logs reads logs directly from disk, so it keeps working even when the daemon itself is unavailable, and vincent task transcript renders a task's transcript as human-readable output, NDJSON or raw agent JSONL, removing the need to dig through Vincent's data directory by hand.

Tasks from GitHub issues, with spending limits

For projects connected to GitHub, Vincent can now create a task from an issue, pulling in the title, description, URL, labels, assignee, milestone and issue number, and filling in workflow fields where the definitions match. The release notes position this as a way to make an issue the starting point of a reproducible workflow rather than a description of work waiting for a person to pick it up.

The release also introduces per-task spending limits, giving autonomous workflows explicit boundaries on how much agent execution they may consume — described in the notes as a safety mechanism for long-running or heavily automated setups.

Workflow authors get two new helpers as well: vincent workflow init scaffolds a starting workflow, and vincent workflow render renders a template without executing it, so generated output can be validated before anything runs. Further changes include idempotency keys for task creation, stricter checks around restricted agent execution, agent adapter version and capability reporting, safer process recovery, workflow source tracking, new built-in workflows, improved retry and polling behaviour, better wrapping of long content in the TUI, and macOS signing and notarization work.

Why it matters

The project's author is explicit that Vincent is not another chat interface around an LLM: the unit of work is a workflow, with agents as one tool alongside commands, checks, control flow, human gates, retries and parallel execution. That framing matches a wider shift in agentic tooling, where the bottleneck is moving from generation quality to operability. Status reporting, notifications, scriptable control, inspectable transcripts and hard budgets are precisely the primitives needed to run agents inside CI or other unattended contexts. The spending limits in particular acknowledge that unsupervised autonomy needs enforceable boundaries rather than good intentions. For teams experimenting with agentic software development, v0.7.0 is a useful signal of which capabilities matter once agents leave the interactive session.