PhasorGate
First things first, we need to focus on permissions and tools. The instructions can be fine-tuned later, but having a shared agent with inappropriate access is the kind of error that's a hassle to correct.
When a development team starts using AI agents, a common problem appears: everyone creates their own configuration. Instructions drift, workflows become inconsistent, and teammates end up repeating the same setup.
One approach is to create a shared agent with a clearly defined role and reusable instructions.
Sharkly.ai supports shared agent configurations through Personal and Space visibility options. An agent configured for a Space can be accessed by members who have access to that Space, allowing teammates to reuse a common setup instead of creating separate copies.
How to approach it
1. Give the agent one responsibility
Start with something specific, such as reviewing code changes, identifying missing tests, or summarizing technical tasks.
2. Standardize the instructions
Define what the agent should check, how it should format its output, and when it should ask for human review.
3. Configure visibility carefully
Choose the appropriate visibility level and verify that the intended teammates can access the relevant resources. Sharing an agent does not automatically grant access to every repository, credential, or execution environment.
4. Reuse the configuration
Keep the core instructions consistent while providing task-specific requirements and context for each assignment.
5. Validate the results
Use tests, code reviews, and clear acceptance criteria to check the agent’s work. Shared instructions can improve consistency, but they do not guarantee correct results.
Where Sharkly.ai fits
Sharkly.ai provides a way for teams to organize and reuse AI agents within a shared workspace. This can be useful when several developers need a common workflow without independently rebuilding the same configuration.
Documentation: https://help.sharkly.ai/docs/agent-and-computer/agents
Question for the community: If you build a shared agent for your team, what would you standardize first: instructions, tools and permissions, or the output format?
First things first, we need to focus on permissions and tools. The instructions can be fine-tuned later, but having a shared agent with inappropriate access is the kind of error that's a hassle to correct.