How Wagglet works
Move the prepared task. Keep every account with its owner.
An author makes the work exact. Another teammate claims it and supervises an agent on their own authenticated subscription. Wagglet keeps delivery and review as separate decisions.
The handoff
Five decisions, each with one owner
- 01
The author prepares two audiences
The task carries detailed LLM instructions for the agent and concise Human instructions for the teammate supervising it. Scope, constraints, repository context, evidence, and stop conditions belong in the task before it is offered to the Team.
- 02
A qualified teammate claims the work
The runner chooses work they may access and can judge. They use their own Claude Code, Codex, or other agent account; Wagglet does not transfer the author's login, provider key, subscription, or token balance.
- 03
Wagglet creates a bounded agent handoff
Copy LLM instructions gives the runner a complete prompt, a read-only claim-rooted context capability, and a fallback snapshot. The agent can refresh permitted current context, but the starting capability cannot deliver or change the task.
- 04
The human and agent run the task together
The agent carries the detailed brief and performs the work. The human supplies authorized access, real-world observation, judgment, and the checks named in the Human instructions. Either participant should stop when authority or evidence is missing.
- 05
Delivery returns to an independent verdict
A separate later prompt lets the same agent session report what happened. That moves the task to Delivered; an authorized reviewer still accepts it, sends it back to the runner, or reopens it. A reported branch, accepted task, merged change, and deployed release remain different facts.
- AuthorPrepareAgent context + human runbook
- RunnerClaimRight access, judgment, and capacity
- PairRunExecution with active supervision
- AgentDeliverEvidence and an honest report
- ReviewerDecideAccept, send back, or reopen
The Dual Prompt
The agent and runner need different instructions
- LLM instructions
- The intended outcome, working context, in-scope and out-of-scope boundaries, constraints, tests, evidence, stop conditions, and delivery format.
- Human instructions
- Why the work matters, how to start, what access or judgment only the runner provides, what to inspect, when to stop, and who decides next.
- Shared responsibility
- The agent keeps the detailed brief active and surfaces uncertainty. The human watches the real environment, corrects course, and performs judgment-heavy checks.
- Independent review
- The participant who performed the work reports it; the authorized reviewer responsible for the outcome decides whether the evidence is enough.
Two agent workflows
Use only the access the job needs
- Task Handoff
- Best when a teammate is running one claimed task. The starting credential is read-only, claim-rooted, and short-lived; delivery arrives through a separate later prompt. Read the guide.
- MCP Workspace Workflow
- Best when one person wants an agent to explore or act across the records currently authorized for them in one Team. Connections are named, revocable, and rechecked on every request. Read the guide.
Use judgment
Some work should not be handed off
- Keep work Human-only when an agent should not receive its instructions or perform its lifecycle.
- Do not give an unqualified runner a high-risk decision merely because an agent can produce an answer.
- Do not place credentials or unnecessary sensitive data in task or human instructions.
- Do not claim work is Done from a plausible report; require the evidence named before the handoff.
- Do not use Wagglet to work around provider account, subscription, or usage rules.
Next steps