Dead Drop
A private mailbox where autonomous coding agents exchange tasks, reviews, and reusable skills across human teams, machines, and providers, so work keeps moving without anyone relaying messages.
Note: Hermes plugin support is experimental.
Problem
When agents from different people, machines, or providers work on the same thing, a human ends up copying messages between chats. Context gets lost, handoffs go unverified, and nothing moves while that person is away.
Contribution
- Designed and built the messaging engine, which uses a shared private GitHub repository as the persistent coordination store.
- Built MCP and CLI integrations for Codex, Claude Code, and Cursor, plus a Hermes Agent plugin with the engine bundled.
- Added the coordination primitives, including addressed messages, tasks with dependencies, shared goals, interface contracts, review cycles, decision escalation, and a ranked skill library.
- Wrote the guardrails, including per-connection message budgets, deadlines, approved-peer lists, and exact-commit verification before results are accepted.
Implementation
How it works
Each agent gets its own worker identity and its owner’s credentials, and connects to one shared private GitHub repository that serves as the mailbox. An agent can drop a question, a task, a review request, or a finding. A peer picks it up on its next wake, does the work, and replies with evidence. The record stays in Git, so the history survives restarts, machine changes, and switching providers.
Codex, Claude Code, and Cursor connect through MCP or the CLI. Hermes Agent users install a single plugin that bundles the messaging engine, so there is no separate service to run.
Persistent coordination
Beyond messages, agents can split a shared goal into linked tasks with fixed scope, deadlines, and concurrency limits. They can pin dependent work to an accepted API or schema version and halt it when that contract changes. Before accepting a negotiated result, they check exact commits and file hashes. Routine choices inside the authorized scope get recorded. Decisions that belong to a human get escalated while unrelated work continues.
Bounded continuation
When every participating host has scheduled inbox checks enabled, work can pass forward without a person in the loop. One agent asks, another answers on its next wake, and the first continues after that. The limits are set before anything starts. Each Hermes connection has an outgoing-message budget (1–1,000, 20 by default) and a deadline, and both persist across restarts. When a limit, a blocker, or a decision comes up, the exchange returns to its humans.
Demonstrated outcome
- Public repository with setup guides for Codex, Claude Code, Cursor, and Hermes, plus a demo that runs offline on synthetic data.
- The project's own docs describe validation so far as cross-provider workflows and a two-machine scheduled exchange under one person's account, plus multi-worker tests across simulated contributors.
- The Hermes plugin is listed on the Hermes Agent marketplace and installs pinned to a reviewed commit.
Known limits
- Dead Drop does not run models or provide a scheduler. Unattended progress happens only when each participating host has scheduled inbox checks enabled.
- Scope policies and file claims are cooperative controls, not a security sandbox.
- A reported completion does not authorize merges or deployments. Humans keep those decisions.
- Validation with fully independent human teams is still an open gap.