Simulator Arena · Submit a lab solution and get it graded here

[akash-coded]

Post a lab solution as a comment here and a bot grades it and replies in the thread. No clone, no Codespace, no local Python — useful on a phone, on a locked-down work laptop, or when you want your reasoning visible to other people.

The format

Two parts: a /lab line, then one python block.

/lab AGL-01

```python
def dispatch(tool_use, registry):
    ...
```

Anything else in the comment is ignored, so explain your reasoning around the code. That is the part other people learn from.

What runs

✅ Public checks	The same ones lab run gives you
✅ Break phase	The failures that end real runs
❌ Hidden checks	Never here — publishing their output would spoil the lab for everyone reading

A green reply is most of the way, not all of it. For the hidden checks, run lab submit <ID> in a Codespace or locally.

Which labs

Any of the ten. Per-lab threads exist for the four starting points below — use those so submissions stay together, or post here for anything else.

Lab	Track	Level
PDL-01 Agent, workflow, or script	product	🟢
AGL-01 Dispatch a tool call	agent-loop	🟢
AGL-02 Close the loop	agent-loop	🟢
AGL-03 Stop the loop	agent-loop	🟡
TOOL-03 Fail honestly: no bare empties	tools	🟡
MEM-01 A buffer that cannot overflow	memory	🟡
RET-05 Citations that survive verification	retrieval	🟡
MAS-02 Cost a topology before you build it	multi-agent	🟡
EVAL-03 A gate that can say no	evaluation	🟡
PROD-02 Failover that cannot be silent	production	🟡

Full catalogue and prerequisites: PATHWAY.md.

How your code is run

It executes on GitHub's runners, so the design assumes submissions are hostile: the grading job has no token at all, and your code runs in a container with no network, 512 MB, 1 CPU, a read-only filesystem and a 45-second kill timeout.

Practical effect: no network, no filesystem, no long-running work. No lab needs any of those.

Full details: ARENA.md.

Etiquette

• Wrap a full solution in <details> so others can still attempt it
• Say what you decided and why — every lab's Learn phase asks for a decision, and arguing about it here is worth more than the code
• A failing submission is a fine thing to post. "Here is my attempt and the check I cannot pass" gets better replies than silence

The drill sequence — nine bites, every track once

Eight to twelve minutes each, in four kinds: implement (fill the TODO), fix (find the planted bug), blank (fill the blanks), predict (say what the code does before running it). Each reply names the skill you demonstrated and points at the next one. Submit in the drill's own thread:

#	Drill	Kind	Level
1	AGL-101 · Pull the answer out of a model message	implement	easy
2	AGL-102 · Two bugs in a working dispatcher	fix	medium
3	TOOL-101 · Fill in the honest empty result	blank	easy
4	RET-101 · Predict what fits in the budget	predict	medium
5	MEM-101 · The summariser that runs every turn	fix	medium
6	EVAL-101 · Which bars are absolute?	blank	medium
7	MAS-101 · Cost this topology by hand	predict	medium
8	PROD-101 · Log the field that makes failover visible	implement	easy
9	PDL-101 · A classifier that promotes by default	fix	easy

After the ninth: your first full lab, /lab AGL-01.

Trainers: /assign @user AGL-101 TOOL-101 --session "Day 3" --due 2026-09-05 in any thread here. Progress: scoreboard · Hands-on Tracker.

Lap two — one level deeper, and the tracks lap one lacked

Ten more, same shape. Start here if lap one is done:

#	Drill	Kind	Level
1	BED-101 · Predict the roles after two tool round trips	predict	easy
2	AGL-103 · Detect a loop that is stuck, not slow	implement	medium
3	TOOL-102 · Catch a hallucinated argument before it crashes the tool	implement	medium
4	MEM-102 · Predict which turns survive the trim	predict	medium
5	RET-102 · A rank fusion that ranks the wrong way	fix	medium
6	MAS-102 · A swarm that does not know how to stop	fix	medium
7	EVAL-102 · Which properties does this golden set violate?	predict	medium
8	PROD-102 · Retryable or fatal — and what an unknown error is	implement	easy
9	PDL-102 · Derive the abstention target from real traffic	blank	easy
10	INT-101 · Predict the A2A task lifecycle	predict	easy

After the tenth: /lab AGL-02 — close the loop, with a Break phase.

Trainers: /assign @user AGL-101 TOOL-101 --session "Day 3" --due 2026-09-05 in any thread here. Progress: scoreboard · Hands-on Tracker.

[akash-coded]

/assign @akash-coded AGL-101 TOOL-101 RET-101 --session "Day 3 · hands-on" --due 2026-09-05

[github-actions[bot]]

📌 Assigned to @akash-coded

_{session Day 3 · hands-on · due 2026-09-05 · set by @akash-coded}

Item	What it is	Time	Submit with
AGL-101	Pull the answer out of a model message · easy	~8 min	/drill AGL-101 + a ```python block
TOOL-101	Fill in the honest empty result · easy	~8 min	/drill TOOL-101 + a ```python block
RET-101	Predict what fits in the budget · medium	~10 min	/drill RET-101 + a ```python block

Post each solution as a comment in this thread (or the item's own Arena thread). The bot grades it and replies; retries are welcome and tracked — a pass after a retry counts.

_{Progress appears on the Hands-on Tracker and in SCOREBOARD.md after the next sync.}

[akash-coded]

/leaderboard

[github-actions[bot]]

🏆 Leaderboard

#	Learner	Passed	After retry	In progress	Attempts
1	@akash-coded	1	0	1	2

Hardest right now: AGL-102 (0% pass, 1 tried) · RET-101 (100% pass, 1 tried)

_{Scoring: a pass after a retry outranks a first-time pass — retrying is the behaviour being rewarded. Full detail on the scoreboard.}

[akash-coded]

/progress

[github-actions[bot]]

📊 @akash-coded

Item	Outcome	Attempts	Best	Next
AGL-101 Pull the answer out of a model message	📌 Assigned	0	—
TOOL-101 Fill in the honest empty result	📌 Assigned	0	—
AGL-102 Two bugs in a working dispatcher	🔁 Attempted	1	4/5
RET-101 Predict what fits in the budget	✅ Passed	1	5/5	/drill MEM-101

Assigned, not yet attempted: AGL-101 (due 2026-09-05), TOOL-101 (due 2026-09-05)