/loop agentic infrastructure

Create, edit, and improve durable agents.

Loop architecture

Goal, create, edit, and improve loops for real agent work.

OpenPond loops are long-running workflows, not disposable prompts. They preserve source, checks, artifacts, and decisions so agents can build, revise, and improve durable systems.

01 Goal loop

Start with the outcome, then let the sandbox own the run.

A goal loop turns a desired result into scoped work, gathers project context, runs inside an isolated sandbox, and reports progress until the work is complete or clearly blocked.

Example: ask for a billing workflow, and the loop tracks the objective, plans the implementation, runs checks, and keeps the status visible as it works.

02 Create loop

Create durable workflows with your integrations already wired in.

Create loops scaffold source-backed agents, workflows, schedules, tools, and integration requirements. They are built for complex automations that need code, validation, credentials, and repeatable runtime behavior.

Example: create a support triage agent that reads Slack, checks a CRM, writes a summary, and ships with validation cases for the expected handoff.

03 Edit loop

Change existing workflows without losing the audit trail.

Edit loops work against an existing sandbox project. They inspect the current source, apply targeted changes, run the project checks, and leave the result as reviewable work instead of a one-off chat mutation.

Example: update an agent to add a SharePoint import path, then verify the source contract, evals, and generated manifest before the change is marked ready.

04 Improve loop

Use real run evidence to make agents better over time.

Improve loops review chat logs, traces, failed checks, user corrections, and completed run artifacts. The loop turns that evidence into source changes, stronger evals, and clearer operating instructions.

Example: after a few weak handoffs, the loop reviews the conversations, finds missing decision criteria, updates the workflow, and adds tests that prevent the regression.

Board

Native issue tracking built into the sandbox.

The board gives agent work a product surface: clear queues, visible run state, reviewable outcomes, and enough context for a person or agent to resume work without reconstructing the thread.

Every request becomes a work item with status, owner, history, and sandbox context.
Backlog, queued, running, needs review, done, and failed states are native to the workspace.
Loop progress, artifacts, checks, and review notes stay attached to the issue instead of disappearing into chat history.

OpenPond sandbox board with work item status columns

Desktop app

Run agent loops from the workspace where context lives.

The desktop app keeps cloud sandboxes, local files, agent conversations, and workflow state close together. It is built for the moments when an agent needs the repo, the logs, and the human review path in one place.

Run the same sandbox loops next to local files, code, chat, and runtime state.
Move between cloud projects and local work without changing the agent contract.
Keep create, edit, and improve work reviewable even when the useful context starts on your machine.

Open source

Check out our repositories.

The core pieces are public, inspectable, and built to be used directly when you want to wire OpenPond into your own agent workflows.

openpond/openpond-agent-sdkThe source-first SDK for authoring loop agents, workflows, evals, traces, and generated runtime artifacts.openpond/openpond-appThe desktop app for running OpenPond loops beside local projects and cloud sandboxes.openpond/openpond-codeThe CLI and TypeScript client for creating sandboxes, running loops, and working with OpenPond from code.