Your Agents Don't Need Smarter Prompts. They Need a Handoff Contract.
Your Agents Don't Need Smarter Prompts. They Need a Handoff Contract.
Every multi-agent setup I have built starts the same way and breaks the same way. One agent researches, one writes, one codes, one publishes. It works on the demo. Then somebody adds a fifth agent, and the pipeline quietly turns into a group chat where every agent is told everything and nobody is told what to hand to whom. Nothing errors. The output just stops being traceable.
The missing piece is never intelligence. It is a contract: a file that says which agents exist, what each one is responsible for, and — the part everybody skips — exactly which agent receives its output next. I went looking for the cheapest possible version of that contract and ended up running AgentDelegator CLI, a $5 Node.js tool that scaffolds it. This post is what I measured when I ran it, from a clean copy, on my machine, today. Including the parts that are honestly thin. It is a product I sell, and this article says so instead of pretending to be a neutral review.
The specific failure it targets
Read the bug reports on any multi-agent pipeline and you find the same three sentences: "which agent ran first?", "did the writer see the researcher's output or the summarizer's?", and "why did the launcher fire before the builder finished?" None of those are model problems. They are ownership problems. A pipeline without declared ownership has no observable order, so when the output is wrong you cannot tell whether a prompt failed or the wiring did.
A handoff contract fixes that by being boring on purpose. Each agent gets an id, a role, a list of capabilities, and a handoffTo list. The graph is data, not tribal knowledge. Once it is data, the wiring is reviewable in a pull request like anything else.
What is actually in the package
Small. I measured every file rather than trusting the listing:
bin/agentdelegator.js— 7,109 bytes, a single CommonJS file. This is the entire program.package.json— 588 bytes. One dependency:commander ^15.0.0.engines: node >=18.README.md— 1,632 bytes, command reference.node_modules/— 207,829 bytes, which iscommander15.0.0 and nothing else.
Three commands, verified with --help: init [workflow-name], status, run. There is no daemon, no account, no API key, and no network call anywhere in the source — I read the whole 7 KB file in one sitting, which is itself the argument for buying something this small.
Disclosure: AgentDelegator CLI is my own product, sold on Gumroad. I earn the sale price if you buy through this link, and I built it for my own agent pipelines. What follows is an account of what the bundle actually does when run — not a third-party review.
Get AgentDelegator CLI — $5 USD (price read from the live listing on 2026-11-22 immediately before writing this, from two independent fields on the page: product:price:amount=5.0 and the JSON-LD offer at 5.0 USD): https://nandminc.gumroad.com/l/wbyhphb
What you get: a Node.js 18+ CLI with three commands — scaffold a workflow directory, print the workflow's agent graph and handoff chain, and a run stub. One npm dependency. macOS, Linux, Windows.
What I measured running it
Clean copy, empty temp directory, Node v26.7.0. These are actual outputs.
1. init scaffolds a real, readable contract
agentdelegator init my-workflow created one directory with three files: agents.json (1,223 bytes), README.md (832 bytes), AGENTS.md (1,634 bytes). The config it writes is a four-agent chain, not an empty skeleton:
researcher → strategist → builder → launcher
Each agent carries id, name, role, capabilities and handoffTo. The researcher declares handoffTo: ["strategist"], the strategist hands to the builder, the builder to the launcher, and the launcher hands to nothing — which is how you can tell it is the terminal node. AGENTS.md documents the field meanings plus a capability vocabulary (web-search, analysis, coding, sales-copy…), so the file is editable by someone who has never read the source.
It refuses to clobber: running init my-flow a second time in the same place exits 1 with Workflow directory 'my-flow' already exists. Small thing, correct thing.
2. status renders the wiring as a human-readable graph
From inside the workflow directory, status prints the workflow name, the creation timestamp, the flow string, and for each agent: its role, its capabilities, and Hands off to. That last line is the whole reason to use the tool. The chain is legible to a reviewer who is not holding the prompts in their head.
3. run is a stub — and says so out loud
This is the honest part, and it is why the price is $5 rather than $50. agentdelegator run prints:
🚀 Workflow execution stub
This is an MVP - actual agent execution would require
integration with an AI orchestration framework.
It exits 0 and writes no state file — I listed the directory before and after; it is unchanged. Nothing is executed, no agent is called, no LLM is invoked. If you read "CLI for autonomous task delegation" and expected it to drive your agents, that expectation is not met by this build. What you are buying is the contract and the scaffold, not a runtime.
Where it breaks — and you should know this before paying
I pushed deliberately broken configs at status to find the edges. They are real, and they matter more than the happy path.
Every validation failure exits 0. This is the one that will bite a CI user. No agents.json at all: it prints "No workflow initialized in current directory" and exits 0. Malformed JSON: it prints the JSON.parse message to stderr and exits 0. An agent with capabilities missing throws a TypeError internally, which the tool catches, prints as an error, and — exits 0. If you gate a pipeline on agentdelegator status, a totally unconfigured directory will read as success. Check the output text, or wrap the call yourself.
handoffTo is never validated against the agent list. I set an agent to hand off to ghost, an id that does not exist. status printed Hands off to: ghost and exited 0. There is no dangling-reference check, no cycle detection (a → b → a renders happily), and duplicate agent ids are accepted silently — my two agents both called a both printed. The contract is real, but it is not enforced. Treat status as a renderer, not a linter.
It does not run without npm install. I copied the bundle without node_modules and the CLI died in cjs/loader on the missing commander require — expected for npm, but worth saying plainly because the source itself has no network code and it is easy to assume it is standalone. Run npm install in the package directory first.
From scratch vs. a JSON file vs. this CLI
| Dimension | Doing it yourself | A hand-written agents.json | AgentDelegator CLI |
|---|---|---|---|
| Getting a contract | Design the schema, then document it | Yours to invent, per repo | init writes a 4-agent chain in ~1 second — measured |
| Reviewability | Depends on whether you wrote docs | Yes, if the fields are agreed | status prints role + capabilities + handoff per agent — measured |
| Validation | You write it | None by default | None — dangling ids, cycles and duplicate ids all pass, and errors exit 0 |
| Executes agents | Your orchestrator | No | No — run is a documented stub |
| Install cost | An afternoon | Minutes | One dependency (commander), Node 18+, npm install |
| Cost | Your time | Free | $5 USD, one-time, no account |
Who it is NOT for
I would rather you skip it than regret it:
- If you want a runtime that actually executes agents — this is not it.
runis a stub that prints an apology and returns 0. For execution you want a framework: I compared CrewAI, LangGraph and AutoGen by hand in AI Agent Frameworks in 2026, and any of them is the right tool for that job. - If you need schema validation or a config linter — there is none. Dangling handoffs, cycles, duplicate ids and missing fields all pass cleanly, and broken configs exit 0. Do not wire
statusinto CI as a gate. - If you already have a working orchestrator config — you are paying $5 for a template you have largely solved. Read the four-agent example above and copy its shape instead.
- If your stack is Python or Go — it is a Node CLI, and
agents.jsonis a data file you could write yourself in two minutes. The value is the agreement on fields, not the language. - If you are shopping by reviews or ratings — there are none to quote. This is a small new listing and I am not going to invent a logo wall or a testimonial to make it look established.
- If you need CI integration out of the box — no GitHub Action, no pre-commit hook. You get a CLI; wiring it up is yours.
Honesty block
The part marketing copy drops:
- Example figures in this article are illustrative planning inputs — not market rates and not earnings.
- Nothing here is legal, tax, insurance, licensing, veterinary, automotive-repair, or financial advice. It is a description of a small code tool.
- I make no claims of CRM, map, API, or invoicing integrations the bundle does not ship — it ships none, and the source contains no HTTP client at all.
- No sales numbers, no customer counts, no testimonials, no benchmarks, no star counts borrowed from other projects. I have none to quote, so none appear.
- Every command output quoted above was produced on 2026-11-22 from a clean copy of the bundle, against fixtures I wrote for the test, in an empty temp directory. Where a path was broken, it says so.
The verdict
Buy this if the thing you are missing is the contract: one file where every agent in your pipeline has a name, a job, and a declared next hop, generated in a second and readable by a reviewer in a minute. Five dollars for a shared vocabulary is a fair trade, and the status output alone ends the "who ran first" argument.
Do not buy it expecting an orchestrator. It does not execute anything, it does not validate its own config, and every error path exits 0 — three facts the listing will not tell you and this article will. If you need execution, read the frameworks comparison first; if you need the wiring to be a reviewed file, this is the cheapest version of that idea I have found.
Questions before you buy: nandminc2020@gmail.com. We answer within 24 hours.
If your next question is what the agent workflow should actually contain once you have the contract, that is this one: How to Build an AI Agent Workflow (Step-by-Step Guide).
Comments
Post a Comment