agents & automation
Work that keeps going
One task at a time is only the start. Orvlet runs a personal agent in the background, fans work out to helpers on different models, repeats until checks pass and starts on a schedule — and every step still goes through your approvals.
Dots: your own primary agent
Dots is one named agent with your goals, standards and voice. Every task it runs starts from them, and it keeps going while Orvlet is open — or on an always-on machine you own. It never runs on a hosted service.
- Several projects in one relationship. It carries ongoing work across your local projects and brings back finished results: pull requests, documents and analyses.
- Learns from feedback. What you say about its work becomes standing preferences it applies next time.
- Proactive, read-only. When you are busy elsewhere it looks for ways to help using read-only tools. It cannot send messages, change app content or control the browser or computer on its own.
- Rules you set. Built-in rules decide when it asks; your own rules allow, ask or block actions; an auto-reviewer checks risky ones; sensitive tasks such as password changes always stay with you.
- Activity view. One place shows what it is doing in the background, with Redirect, Pause and Stop.
Teams of agents and models
Split big jobs into parts, run them in parallel and bring the results back together. Helpers inherit the parent task's permission mode, never more.
- SubagentsDelegate with an @mention or the task tool; give each helper its own model, prompt and tools. Local helpers are admitted only if memory allows.
- Dynamic workflowsThe model writes a fan-out plan, Orvlet runs the steps in parallel and shows a graph and timeline of every branch.
- Worktree isolationHelpers can work in their own Git worktrees and return typed results, so parallel edits never collide.
- /researchA planner, one researcher per question and a writer produce a cited report; code — not the model — enforces source triangulation.
- /reviewA structured code review with P0–P3 findings and a verdict, or presets for security, tests and style.
- Consensus votingThree different models each review a diff and vote accept, revise or reject; ties go to the stricter vote.
- Mixture of agents
/moaasks two to five proposer models and has an aggregator combine their best answers. - Advisor and lead/workerPair a fast model that does the work with a stronger one it consults, or switch models per step.
- /boost and /teamworkThrow more models at a hard prompt, or coordinate teammates that message each other inside one project.
Automation that runs on its own clock
Schedules and triggers start ordinary tasks — with pre-approved permissions you choose in advance, and results waiting in your inbox.
Scheduled tasks
Cron-style automations and templates; the agent can create its own schedules when you ask it to. Orvlet can keep the Mac awake while tasks run and warns before you quit with work in progress.
Goals and loops
/goal keeps a persistent
objective across turns, /loop repeats a step until a check passes, and
heartbeats revisit an existing thread on a timer. Doom-loop detection stops a task that
keeps repeating itself.
Event triggers
Watchers react to file changes or a new commit on main; an incoming webhook with token authentication can queue a task; MCP event subscriptions can start one too.
Pull requests on autopilot
Watch pull requests and fix failing checks, draft reviews when a pull request opens, route CI failures and review comments back into the task, and merge with the method you choose.
Security scans on a schedule
Scan a whole repository on a timer or on new commits, keep findings across runs, report only new ones and prepare fixes for review.
Headless and scripted
Run any of it from the
orvlet CLI or a script, with JSON output, a structured output schema and
a restricted mode for evaluation runs.
Background work, foreground control: Redirect, Pause and Stop are always one click away.
Photograph — Gabriele Garanzelli via Unsplash
Memory, plans and recovery
Long work needs a good memory and a way back.
- Layered memoryProject knowledge and personal memories with consolidation, time-aware review and corrections — and a per-task switch to turn it off.
- Methods and skillsWorkflows the agent learned become reusable methods; skills improve from what worked, with you reviewing the change.
- Plan modeThe agent proposes a plan with inline comments; nothing is changed until you approve it.
- Checkpoints and rewindUndo a turn's file changes, fork from an earlier message, or replay a session from a checkpoint.
- Context that lastsAutomatic compaction at a threshold you set, a context meter and a breakdown of exactly what the model sees.
- Kanban and inboxTasks that need you surface in one place, and a board tracks work from pull request, CI and review facts.
Costs you can see
Every task records tokens and, where a provider publishes prices, an estimated cost per turn and per task. Budgets, provider-reported rate limits and a request inspector that captures the exact payload keep surprises out — and local models cost nothing per token.
Still asks before it acts
Automation does not widen permissions. Scheduled and background tasks use the permission profile you gave them, helpers never exceed their parent, and anything outside those rules waits for you.