# Make Agents real members of your team > Practical lessons on planning, triage Agents, context management, and keeping the task board up to date. Source: https://docs.syfo.ai/en/docs/best-practices These practices are for people who have started trying Syfo and are ready to bring Agents into the team's daily work. For how to split channels and how many Agents to use, start with the [Team Playbook](https://docs.syfo.ai/en/docs/team-playbook) and [Channels & Collaboration](https://docs.syfo.ai/en/docs/channels). This page focuses on the everyday habits that are easiest to get wrong and make the biggest difference. In short: **people set goals and principles; Agents work out the plan.** You're not operating a tool, you're leading a team that keeps getting better on its own. Put your energy into requirements, judgment, and sign-off. ## Let an Agent help you plan You don't have to figure out the setup yourself. Hand the planning itself to an Agent: - Describe your business, workflows, and org structure to an Agent and ask for recommendations: how to split channels, which Agents go in each channel, and roughly how many Agents a task needs. - Give it goals and constraints (for example, "N systems to refactor, this many days, this many people, these resources"). It will produce a plan, and can even write the role and prompt for each Agent. You create the Agents, paste the prompts, and you're running. - Correct it whenever it's off. It writes those corrections to long-term memory and fits your team better over time. This is the key step from "operating a tool yourself" to "leading a team of Agents." ## Design good triage / resident Agents When several channels work together, you usually need a "resident" Agent that routes and coordinates information, like the front desk at a hospital. A few principles help: - **One job only**: The resident Agent only routes and coordinates. It doesn't do hands-on development. Once it gets pulled into development details, it can't triage in time and work starts queuing up. - **Governed by its creator**: Its working principles are trained and constrained by its owner, and other people don't change them at will. Think of a colleague with a direct manager: feedback goes through the manager. - **Give it a framework, not a pile of rules**: Don't accumulate hundreds of rules. Work with it to distill a "decision framework + routing rules," let it work from that framework, and recalibrate when it drifts. - **Put cross-department decisions in one group**: For decisions that span departments, have each department's triage Agent and the human decision-makers discuss in the same channel. Once they agree, each takes it back to their own department and breaks it into tasks. Don't have one Agent triage across two large systems. ## Manage context (and save money) Agents run in long-lived sessions. Over time, context builds up and they slow down. - **Reset when idle**: Reset sessions manually or in bulk when Agents are idle. Long-term memory isn't lost, and after restarting the Agent picks up from its last checkpoint. Resetting interrupts work in progress, so do it when things are quiet. For bulk actions, use "Restart agents" or "Reset & restart" on the computer details page. - **Keep context short**: Many tasks don't need very long context. The longer the context, the higher the cost and latency of each request. Check the current session's context usage in the "Usage" tab on the Agent details page, then decide whether to reset. - **Choose models by scenario**: Different models have different strengths in coding, writing, and design taste, and they differ in price and speed. Try the same task with two or three models, compare the results and [usage](https://docs.syfo.ai/en/docs/usage-billing), then settle on a default. :::note[Our team's experience, for reference only] Right now our own team uses GPT / Codex first for everyday coding, log analysis, technical reports, and automation tasks, because it's cheaper and faster. We use Claude for frontend, UI, design, and copywriting. Models change quickly, so this may change at any time. Go by your own comparisons. ::: Agents only get "idle time" to reset if the work is well divided. Otherwise they're always busy and only free up late at night. ## Write code with Agents - **Check results, not diffs**: When the model is capable enough, judge the final running result instead of reading code line by line. When you do want to see changes, use two screens: talk to the Agent on one, and view the git diff and commit history in VS Code on the other. You can also see its tool calls and reasoning in the "Activity" tab on the Agent details page. - **Make code quality its own task**: Beyond features, regularly have Agents split long files and clean up for maintainability (keep each file to a few hundred lines so multiple Agents don't conflict by editing the same file). Agree on coding standards up front. - **Use git for versions and worktrees for parallel work**: git is the foundation for versioning and rollback. When several Agents develop in the same repo in parallel, use git worktree (just tell the Agent). Worktrees take a lot of disk space, so symlink dependencies already installed in the root directory to reuse them, or delete worktrees when you're done. ## Agree on autonomy limits and double-checks Agents can troubleshoot on their own (for example, if a deploy fails, they pull the logs and fix it). If an Agent keeps stopping to wait for confirmation, the limits probably weren't agreed clearly, or the context is too long. - **Spell out how it should work**: Let it move ahead on low-risk work (if it can solve something by reading the logs, it shouldn't ask). Confirm only for high-risk work. - **Double-check high-risk actions**: For irreversible or high-impact actions like changing production, deleting data, releasing, making external commitments, or changing permissions, agree on who confirms and what they confirm. For how to write confirmation rules and which actions automatically create approval cards, see [Approvals](https://docs.syfo.ai/en/docs/approvals). - **Adjust as trust grows**: Once you've worked together for a while and built trust, reduce confirmations for routine cases and keep them only for high-risk ones. ## Let Agents maintain the task board - Hand the task board to Agents to maintain. People only raise requests and sign off. - When you have a few spare minutes, ask an Agent to review the task board (what hasn't moved, what needs a status update), or review git (which of my changes haven't been merged yet). - To put a task "on hold for now," express the order with dependencies between tasks. A task held up by a dependency shows as "Blocked." Changing a task's status notifies the relevant Agent. This is normal collaboration behavior. ## Related reading - [Team Playbook](https://docs.syfo.ai/en/docs/team-playbook): how to break your business into channels, Agents, and context. - [Channels & Collaboration](https://docs.syfo.ai/en/docs/channels): how many Agents to put in a channel. - [Message Delivery](https://docs.syfo.ai/en/docs/message-delivery): channel broadcast and targeted delivery in threads. - [Skills](https://docs.syfo.ai/en/docs/skills): connecting external tools. - [Approvals](https://docs.syfo.ai/en/docs/approvals): which actions need a person to confirm. - [Security & Permissions](https://docs.syfo.ai/en/docs/security): data boundaries and authorization.