Your First Orchestration¶
This walkthrough takes you from an empty directory to a working Python module with tests, passing every agent through the full pipeline. You'll create a Fibonacci module and its pytest suite, exercising all 5 agents in sequence — Riko, Senku, Loid, Lawliet, and Alphonse. This takes 2-5 minutes and consumes tokens.
Before you start¶
claudeCLI installed and configured (see installation.md)- Python 3.9 or newer available (
python3 --version) pytestavailable, or willing to install it (pip install pytest/uv add --dev pytest)- Agent Flow plugin already installed (see installation.md)
What you'll build¶
By the end you'll have a fibonacci.py file with an iterative fib(n) function that raises ValueError for negative inputs. Alongside it, tests/test_fibonacci.py will contain four cases covering the base cases, a larger value, and the error path. Both files are created and verified entirely by the agent pipeline — you only provide the seed prompt.
Set up an empty workspace¶
Start Claude Code inside the fresh directory. There are no source files here yet, which is intentional — Riko will confirm the blank slate before Senku plans.
Run the orchestration¶
Once the session is open, run:
/orchestrate Create a Python fibonacci.py module with an iterative fib(n) function and a pytest test file in tests/ that covers fib(0), fib(1), fib(10), and a negative-input ValueError case.
Then sit back and watch the agents work.
What you'll see, agent by agent¶
Riko explores¶
Riko opens by surveying the workspace. Because the directory is empty, Riko will report exactly that — no existing files, no dependencies, no test framework config detected. This is useful signal: Senku will know there is nothing to preserve or integrate with, and will plan a greenfield implementation. You should see Riko's findings appear as a short summary in chat, typically a sentence or two noting the absence of source files and confirming Python as the target language.
Senku plans¶
Senku takes Riko's findings and translates them into a concrete task list. You should see Senku's TodoWrite list appear with items like "create fibonacci.py with iterative fib(n)", "raise ValueError for n < 0", and "create tests/test_fibonacci.py with four test cases". Senku may also note any assumptions — for instance, that fib(0) returns 0 and fib(1) returns 1. If anything in the plan looks wrong, this is the right moment to notice it, though you don't need to intervene.
Loid implements¶
Loid works through Senku's task list one item at a time, writing fibonacci.py first and then tests/test_fibonacci.py. You'll see file-write events appear in the tool call stream. Loid follows the plan closely, so the iterative algorithm and the ValueError guard should both appear exactly as Senku specified. After writing, Loid may run a quick sanity check to confirm the files exist and import cleanly before handing off.
Lawliet reviews¶
Lawliet reads the newly created files and checks for code quality issues — things like missing docstrings, edge-case gaps, or deviations from idiomatic Python. You'll see a verdict line in Lawliet's response, either APPROVED or NEEDS_CHANGES. For a simple Fibonacci module, Lawliet typically approves on the first pass. If Lawliet raises a NEEDS_CHANGES verdict, the feedback loops back to Loid automatically; you don't need to do anything.
Alphonse verifies¶
Alphonse runs the actual test suite and reports concrete output. You'll see the pytest command execute and its output appear verbatim in chat — something along the lines of 4 passed in 0.12s. Alphonse also checks for import errors or collection failures. If everything passes, you'll see the orchestration-complete marker:
Verify the result yourself¶
After the session, confirm the files landed and the tests still pass outside Claude:
Expected output shape: 4 passed in X.XXs (the exact count may vary slightly if Loid splits edge cases into additional tests).
If something goes wrong¶
Agent stalls or session is interrupted — orchestration state lives in .claude/orchestration.local.md inside fib-demo/. Re-running /orchestrate with the same prompt will resume from where it left off rather than starting over.
pytest not found — run pip install pytest or, if you're using uv, uv add --dev pytest, then re-invoke the orchestration.
Verification fails on the first Alphonse pass — this is normal behaviour, not a bug. Lawliet's feedback is fed back to Loid, which corrects the issue and re-submits. The loop resolves automatically within one or two iterations for straightforward cases like this one.
Want to start completely clean — exit the Claude session, then:
Re-run the workspace setup steps to start fresh.