Give each agent one job, keep one shared task list, and let a single agent send anything that leaves your hands. That follows the five plain rules in Tokenrip's guide: one job per agent, claim work before starting it, lock anything that sends, dedupe on one key, and keep one source of truth instead of copies. Tokenrip sells a shared workspace for agents, so treat the rules as a starting point and test them against your own setup.
Quick version
- Add a second agent only after one agent hits a real wall. JetBrains says to add agents only for context-window saturation, a missing capability, or unacceptable sequential latency. JetBrains sells AI agent products.
- Give each agent one job and tell it that it is a helper, not the lead. Microsoft's guidance explains why.
- Keep one task list with an owner, a status and a time limit on every claim.
- Check each agent's output before the next agent uses it.
- Let one agent talk to you and one agent send. You review in between.
Why more than one agent gets messy
Agents do not share a brain. Claude, ChatGPT, Codex and Grok each keep their own memory, and none of them reads the others'. If you run two of them on related work, each one works from a different picture of the world.
That produces two failures worth naming. JetBrains describes synchronization failures, where an agent acts on stale or incomplete state. It also describes cascading failures, where one agent's bad output poisons the agents downstream. Both are what you see when agent B builds on a mistake agent A made an hour ago.
The problem is common at larger scale too. Lyzr says 94% of enterprises report that agent sprawl is already creating security and operational headaches. Lyzr also says only 5% of enterprise agents ever reach production, and that the dropout is mostly at orchestration boundaries, not at agent quality. Lyzr sells agent orchestration platforms, so read that as a vendor's claim. Still, it points at the right place to look: the handoffs between agents, not the agents themselves.
Step 1: check that you need more than one agent
More agents means more to read, more to verify and more that can go wrong. JetBrains advises adding agents only when you hit a concrete wall. If one agent can finish the job in your hour, stop there.
When you do need a team, keep it small. Addy Osmani, writing about coding agents, says three to five teammates is the sweet spot, that token costs scale linearly, and that three focused teammates consistently outperform five scattered ones. He also writes that three focused agents consistently outperform one generalist agent working three times as long. His focus is software work, so the numbers may not carry over to every task.
Our view: for someone with a day job and one spare hour, the limit is not the token bill. It is how much output you can read and judge before bed. Start with the smallest team that clears the wall.
Step 2: one job per agent, stated plainly
Overlap is where duplicate work comes from. Microsoft's multi-agent guidance for Copilot Studio says that if two subagents search the same knowledge base, one finds the answer first, and the second either returns duplicate results or skips its search, adding no value. Give each agent its own sources and its own slice of the work.
Then say so in the instructions. Microsoft notes that subagents do not inherently know they are part of an orchestration, and without explicit guidance they behave as standalone agents and message the user directly. The page is about Microsoft's own product, but the habit applies anywhere: write down each agent's role, and write down that it reports to the lead, not to you.
Step 3: keep one shared task list
A single list replaces the copies that drift apart. Tokenrip's suggested Tasks table has these columns: what, which job it belongs to, owner, status, claimed until, and a link to the result. A spreadsheet can hold that.
The working routine is short:
- An agent picks a task, marks it in progress with its own name on it, and only then starts.
- Every claim has an expiry. A task held by an agent that crashed or ran out of context goes back on the list.
- The agent writes the result link into the row when done.
Pick one key to match records on, too. For companies, Tokenrip says the key is usually the email domain, not the company name, because 'Acme', 'Acme Inc.' and 'ACME Corp' are one company to you and three to an agent.
Step 4: handle dependencies with a pattern
A dependency just means one piece of work needs another to finish first. Developers Digest, updated October 9, 2026, covers six coordination patterns: fan-out/fan-in, pipeline, hierarchical delegation, blackboard, handoff chain, and consensus. Three are enough for most people to start.
Our view: a pipeline is the easiest to follow when you are the only human involved, because you can read the work in the order it happened. Use fan-out only when the subtasks truly do not depend on each other.
Whatever you pick, put a check between the steps. In a pipeline, if stage 2 produces malformed output, stage 3 crashes, so every pipeline needs validation between stages. JetBrains calls this a validation gate: each output is checked before the next agent acts on it. A gate can be as simple as a rule that the output must be in a set format, or a quick read by you.
Step 5: control the outputs
The risky outputs are the ones that reach other people or spend money. Treat them differently from internal notes.
- One voice to you. Microsoft says only one agent should talk to the user per turn, and the parent agent is the one that should deliver the final response.
- Three roles for anything that leaves. Tokenrip suggests one agent prepares, one reviews in batches, and only one sends.
- A lock on sending. Anything that sends, posts or charges needs a lock: one row saying which agent is sending, until a stated time.
- A log written during the send. Before sending, the agent looks the contact up in the log and stops if it is there. If not, it writes the domain to the log as part of the send, not afterwards.
For coding agents the same idea has a name. File locking prevents two teammates from editing the same file simultaneously.
Some work should not go to an agent at all. Our pages on what to never automate with an AI agent and AI agent spending limits cover where to draw that line.
Step 6: verify, because you are the last gate
Speed is not the hard part anymore. Osmani writes that the bottleneck is no longer generation but verification: agents produce impressive output quickly, and knowing with confidence that it is correct is the hard part. With several agents, there is more output to verify, so plan your hour around review, not around starting more work.
Keep each task small. Osmani describes the Ralph Loop, which resets the agent's context each iteration so it avoids accumulating confusion, and says small bounded tasks produce cleaner code with fewer hallucinations than one enormous prompt. The same source says LLM-generated AGENTS.md context files offer no benefit and can marginally reduce success rates, while developer-written context files provide a modest improvement. In plain terms: write your own instructions, and keep them short.
Common problems and fixes
| Problem | Fix |
|---|---|
| Two agents do the same task | Claim work before starting, with a name on the task |
| An agent acts on old information | Point every agent at one source of truth, not copies. Stale state is what causes synchronization failures |
| One bad output spreads | Add a validation gate between agents |
| A helper agent messages you directly | Tell subagents they are subagents |
| A task is stuck on a dead agent | Give every claim a time limit |
| Duplicate emails to one company | Dedupe on the email domain |
What to do this week
- Write the job of each agent in one sentence. If two sentences overlap, merge or cut one.
- Make the task list with the six columns above.
- Choose a pattern and add one check between steps.
- Decide which single agent is allowed to send, and which of your own evenings are for review.
These ideas come from vendor pages, a developer's blog and product documentation, not from controlled studies. Several sources sell agent software. Our view: the common thread is dull but useful. Fewer agents, clear boundaries, one list, and a human at the end.
If you want the habit that goes with it, Sunday Night Truths is a free weekly email. Signing up also sends The Agent Leash, a one-page checklist of 15 checks before your AI agent touches your money. Get Sunday Night Truths.
For how this fits around a job, see how to ship side projects with a full-time job.
