Programs for recurring agent work, with run reports
Oct 18, 2026 Models
A dispatch says when the executor's model cannot run
entry 59 of 59since April 2026
what's new
What's new in ConvOps?
The ConvOps changelog lists every change a user can notice in ConvOps, the operations layer for AI agents, dated by the day it reached the product. Entries run newest first and are grouped by month. Each one says what you can now do, in plain language, and links to the page that explains it.
Define the numbers your agents must report and bind each one anywhere from the workspace down to one step, with a target on the binding and a condition for when it is due. A required metric holds that step until it is recorded, with optional evidence, and the server stamps who recorded each value and where. ConvOps also records the cost, duration and outcome of every run itself. Each workflow template change leaves a marker on the trend of each metric bound to it, so you can see whether a change made runs better.
A week and month board shows what your agents execute and when, next to a tray of work with no day yet and a capacity figure from the last four weeks of finished work. Spread a batch over the week in one step, after a dry run, and a planned task with an executor can start on its day. Each schedule now keeps its own time zone and shows its next five fires before you save it.
Group the schedules behind a recurring job into a program with one owner, purpose and goal, and see at a glance whether it is healthy, late, silent or failing. Agents file a run report with the outcome, what changed, what comes next and any blockers, or a report type your team defines with its own fields.
When an environment gives the agent full git access, the git guidance in every run now tells the agent to fetch, merge and push its own work to the task branch or the chosen branch, never force-pushing. The environment page shows that guidance.
Dispatching a task that is blocked, paused, completed or cancelled is now refused with the reason before anything starts. A stalled task is picked up again on dispatch.
A dispatch is refused, with the fix, when the executor has no model of its own or its model is not usable on any of its provider connections. OpenCode runs and executor checks now get a models file built from your registry, so they see only the models that count.
Each agent names a tier, such as strong, balanced or fast, and a tier sets one model per runtime. Provider connections hold your keys and check which models really answer; executors link connections and hold no keys. Every run records which model each agent used and why.
An isolated run now gets its environment's team, or every enabled agent, plus each agent a workflow step names. The task previews that set before you dispatch, and a repository agent with the same name as a team or step agent blocks the task with the file to rename.
The new Models page lists every model your workspace can use, kept in step with the models.dev catalog every day, plus models you add by hand and an allowlist per workspace. Agents can pin a model per runtime.
Creating an isolated executor, or changing its connections or its agent version, starts a short check that proves its runtime starts and its own model answers. The executor page shows the check while it runs and the result when it ends.
A workflow step can name the catalog agent that does it. An unknown or disabled agent is refused when the step is saved, an agent that an enabled workflow names cannot be disabled, and workflows can be filtered by the agents they name.
The convops command line imports .claude/agents and .opencode/agents files into the agent catalog and exports the catalog back as files. Both directions have a preview that writes nothing.
Retry a task's latest run from the runs panel, the list, or one tool call. Resume on the same working copy or start fresh; the run shows which modes it allows right now.
Keep every agent of a workspace in one catalog. Every save is a numbered revision you can roll back to, and teams group agents with each member pinned to a revision or following the active one. The app has a new Agents page with a Teams tab.
An environment decides who does git with the remote: by default the agent commits and ConvOps pushes, or with full access the agent holds the matched key itself. ConvOps can generate SSH keys for git access, and an environment check proves read and write per repository.
A task or a schedule can name its own environment instead of the project default. The task page shows which environment applies and where it came from, with a picker to change or reset it.
Open any workflow as a graph of its steps, with sub-workflows opened in place. The installed workflows list now shows step counts and how often each one runs.
The dashboard, goals and activity pages now open with charts. Goals show their health and period at a glance, and activity shows events per day, by action and by the busiest actors above the timeline.
When a decision invokes a sub-workflow, the task page shows the parent workflow with the open sub-workflow nested inside it. You see exactly which step the run is on, at every level.
Memories and routes are now filed under a project from your workspace catalog. Recall, context lookups and memory lists can be narrowed to one project, and a memory can be moved to another project when you edit it.
Merge one project into another and its work moves with it. A task that names an unknown or archived project is now refused with a clear message instead of creating a new project by accident.
An environment can set the name and email on every commit a run makes. You also choose which repository receives the work, and whether a finished run lands on a task branch for review or straight on the branch you configure.
Add MCP servers to your workspace and attach them to an environment, with rules that allow or deny each tool. Isolated runs start those servers next to the agent, and the agent is offered only the tools the rules let through.
Each environment lists the ConvOps tools an unattended run is offered, and a person switches each one on or off. Task tools reach only the run's own task unless you widen them.
An executor can now start every run as its own Kubernetes Job: a fresh pod with its own volume and secret, and a time limit. Each run keeps a history with its event log and transcript, and a Stop button.
Prepare once what every run needs: repositories, git access, secrets and variables. A check clones every repository and resolves every secret in a throwaway copy before any run starts, and reports what is ready or failed.
Browse, filter and read your workspace memories and routes in the web app, and add or correct them from the same page. When a new memory replaces older ones, the app tells you which.
Mark one workflow as the router of your workspace. Every new task, bug, initiative or plan starts there, and the router sends it to the right workflow. The app shows a Router badge on that workflow.
Every dispatched run now shows its status, timeline and transcript on the task page, and a new executions page lists runs across the workspace. Your AI client can read the same status and transcript, and stop a run.
Your AI can change one passage of a task description or a step instruction by anchoring on the text around it, instead of sending the whole thing back. A version check stops two agents from overwriting each other, and replacing a long text outright needs an explicit confirm.
Mark when work is committed: now, next or later. Children inherit the horizon of their nearest ancestor, the dashboard groups work by horizon, and four built-in pools list what is committed now, what is actionable, what is next and what has gone quiet.
A workspace setting stops anyone, person or agent, from marking a task done while its workflow is still running. The refusal names the current step, so the work finishes through the workflow. Cancelling stays possible.
Every installable workflow template and pack is now browsable on convops.app, with its real steps, gates and decisions. Pick one, then install it into your workspace in the app.
Each workspace now keeps one list of projects. Tasks, routing rules and every project picker in the app read from it, and a project can be archived when the work is done.
Every change now records the server-verified person behind it next to the label the agent gives itself. The app shows the person on the task and next to the agent on every activity row, so you can always answer who approved that.
A step can name the executor that should run it, and a task can carry its own default. When a run reaches a step meant for a different executor, the agent is told to hand it over. Nothing is dispatched behind your back.
Define executors for your workspace, each naming which AI client runs the work and where. Dispatch a task to one of them, then follow its status and read its transcript from your AI client.
Save a query as a pool, and an agent can ask it what to work on next. The pool is resolved live on every call and returns exactly one task or nothing, so empty is an answer the workflow can branch on.
Scheduled runs that go quiet are picked up and resumed automatically. After repeated tries the task moves to a new Stalled status with the reason, and a parent waiting on it continues instead of hanging.
Create a workspace, accept or decline an invitation and manage your API keys straight from your AI client. Creating or revoking a key still needs a signed-in person.
Opening a task now brings the memories linked to it and similar past tasks, in your AI client and on the task page. One context lookup returns memories, routes and related tasks together.
Pass a short context note with a task or workflow call and it is stored as a memory linked to that task. A step can require that note before it advances, and automatic capture is available as a workspace setting, off by default.
Thirteen new starter workflows in four packs: Content Studio, Sales Outreach, Business Ops and Chat Essentials. They are written for a plain AI chat, with approvals only where something is published, sent or decided.
A single connection now reaches every workspace you belong to. Writes name the workspace they go to, and search and recall look across all of them at once.
Attach a schedule to an existing task and each fire resumes its workflow where it stopped, instead of creating a new task every time. Schedules that spawn a fresh task per fire work as before.
Point your workspace at your own OpenCode server and tasks can run there without anyone typing. The connection is validated when you save it, from a new Execution page in settings.
Your workspace gets shared memory. Agents store what they learn as memories, record where things live as routes, and look both up before they act, so the next run starts from what the last one found.
Write a rule once, globally or for one workflow, and it is delivered inside the step the agent is working on. Edit it and every run picks it up. Policies inform the agent, gates enforce.
Ask your AI client to add, remove, move or update a single step, or to change a workflow's details. Each edit is validated against the whole workflow before it is saved.
Install a whole pack of workflows in one go. Shared fragments come along automatically, and the app warns you before you uninstall something another workflow depends on.
A new Explore page lists every starter workflow with suggested prompts. Install one into your workspace, where it sits next to the workflows you write yourself, and pick your first workflows during onboarding.
Ready-made workflows for goals, plans, phases, features, bug fixes and tasks, with gates between the stages. Every type of work can start on a process from day one.
ConvOps ships continuously, not in big releases. A change goes live as soon as it reaches the product, and this page lists the ones a user can notice, each dated by the day it landed. Some days carry three entries and some weeks carry none, so the dates show the real pace of the work rather than a release calendar.
What counts as a change in the ConvOps changelog?
A change counts when a ConvOps user can notice it: a new feature, a new option or a change in how something behaves. Internal work is left out, such as tests, build pipelines, refactors and infrastructure. Each entry says in plain language what you can now do, not how it was built, and most entries link to the page that explains the feature.
How are changelog entries dated?
Each ConvOps changelog entry carries the day the change reached the product, not the day it was written up. Entries run newest first and are grouped by month. The top of the page shows two dates: the newest entry's day, and the day this page last changed. In the RSS feed, each item is dated at noon UTC on that day, so no reader's time zone moves it to the day before.
Is there an RSS feed?
Yes. The ConvOps changelog has an RSS 2.0 feed at https://convops.app/changelog.xml, built from the same data as this page, so the two never disagree. Each entry becomes one item, newest first, with its title, its area as the category, its full text and a link to the entry's anchor on this page. Add the address to any feed reader to follow every change.
How do I follow one part of ConvOps?
Filter the ConvOps changelog by area. The chips above the list cover workflows, shared memory, runs, agents, models, environments, governance, workspaces and the app, each with its entry count, and picking one leaves only its entries on the page. The filter works on this page only: the RSS feed carries every entry, with the area in each item's category, for readers that sort by category.
Do I need to install an update?
No, not for hosted ConvOps at mcp.convops.app. A change listed here is live for every connected AI client and every new run as soon as it reaches the product, with nothing to install or download on your side. Your AI client keeps its own model, subscription and update cycle, which are separate from ConvOps and from this changelog.
Where can I read more about one change?
Follow the link on the entry. Most ConvOps changelog entries end with a link to the page that explains the feature, such as a product page, a glossary definition or a guide. Every entry also has its own anchor on this page, so you can share a link to one change, and that same anchor is the link in the entry's RSS item.