[FAQ] Help with agent speed of designing and coding

[alexisdacquay]

Course

machine-learning-zoomcamp

Question

How to reduce time taken by the PM, Engineer, QA agent? Everything is so slow.

Answer

• Make sure your tasks are not too numerous. Group them logically (or ask the AI to). 2 is too few. 99 is too many and might get stuck to infinity
• While the Session 1 of the course focussed on a robust process, it might be overkill of simple projects. Prevent bloat and complexity by explicitly asking the AI to review against those criteria
• Ask your AI to review the tasks and issues, it might spot some logical loops, or anomalies in the process. Correcting those before starting can save your hours.
• Specifically ask the orchestrator to be a higher level LLM, while the doers are lower-tiers of LLM
• Give the AI orchestrator agent some monitoring and time-control guidance
• Ask the engineer (coding agent) to follow the lazy ponytail dev principles (https://github.com/DietrichGebert/ponytail)
• Ask your AI to have a dry run and to monitor closely what the sub-agents are doing, it might surface further anomalies
• Once all the above "pass" then you may run the whole process with more confidence.

Checklist

• I have searched existing FAQs and this question is not already answered
• The answer provides accurate, helpful information
• I have included any relevant code examples or links

[github-actions]

✅ FAQ NEW proposal created in PR #395

Please review and approve the changes.

[akufunalzm-data]

The PM, Engineer and QA agents can take longer when a task is too broad, the project is large, or each agent repeatedly reviews the full codebase and conversation history. Running every step sequentially, requesting lengthy explanations and allowing repeated revisions can also slow things down.

A few adjustments may help:

Keep tasks small
Assign one feature, component or issue at a time instead of asking the agents to build the entire application at once.

Give each agent a clear role

PM: Define requirements, priorities and acceptance criteria.
Engineer: Implement the approved requirements.
QA: Test the implementation against the acceptance criteria.

Provide complete requirements upfront
Include the expected functionality, technical constraints, relevant files and a clear definition of completion.

Ask for concise outputs
Request only the plan, affected files and acceptance criteria where detailed explanations are unnecessary.

Limit the working context
Direct agents to review only the files relevant to the current task.

Run independent tasks in parallel
Unrelated activities can run at the same time, while dependent tasks should remain sequential.

Focus QA testing
Test the changed functionality and related risks first, then run the full test suite during final validation.

Set clear stopping conditions
Limit revision rounds and ask agents to report unresolved issues instead of continuing indefinitely.

Reuse established decisions
Record approved architecture, standards, and dependencies so agents don't reconsider them for every task.

Check the execution logs
If small tasks are still slow, look for delays caused by installations, large repositories, test execution, external services, platform load or agent configuration.

A practical workflow is to have the PM define measurable acceptance criteria, the Engineer implement only that scope, and QA test against those criteria. If the process remains slow, please share the prompts, execution logs, and the approximate time taken by each agent. This should help identify the actual bottleneck.