// field notes

Real runs.
Real signals.

These are usage notes from active ShrimpHub controller runs. The numbers come from preserved controller history and agent logs; they show scale, validation loops, recovery patterns, and what the tool is already useful for.

25k+
agents run
106
projects tracked
21.1B
tokens recorded
348k
calls in the last 7 days

// case note

One preserved run produced a playable Godot game milestone.

Chess-2 was built from a design document by an unattended agent swarm. It reached a playable milestone, and the case study includes sanitized broadcast excerpts, the self-healing bug chain, the agent-written knowledge base, and the final QA finding.

41agents
14h21 minutes
123tests
41commits

// recent activity

Five days.
11B tokens.

A recent preserved window shows sustained multi-project throughput. This was not a benchmark; it was normal use during active development of the controller, prompts, and task types.

May 25
1.66B
May 26
1.49B
May 27
2.97B
May 28
2.77B
May 29
2.11B

// peak day

3.0B

Recorded tokens in a single day — May 27 — across projects running in parallel.

May 29 is a partial day at time of writing. These are not bursts — the system runs continuously, unattended, across projects with different task queues and priorities.

// agent workload

22,759 agent runs.
Here's the shape.

Across all recorded runs, agent workloads vary significantly. The long tail is where the real work happens.

LLM turns per agent run
Median 17 turns
Average 31.6 turns
90th percentile 72 turns
95th percentile 110 turns
99th percentile 200 turns
Max observed 200 turns
Tokens per agent run
Median 745k tokens
Average 1.50M tokens
90th percentile 3.41M tokens
95th percentile 5.60M tokens
99th percentile 13.33M tokens
Max observed 33.9M tokens
Reliability events
Tasks completed 9,929
Tasks failed 1,502
Rate-limit events handled 2,986
API retries 13,866
Web searches 83
Agent loops recorded 296k
Log corpus
Agent runs recorded 22,759
Agent log files on disk 5,791
Logs with LLM activity 5,787
Explicit LLM calls 141,693
LLM activity events 308,778

Across 22,000+ recorded agent runs, the failures looked like real engineering failures: bad fixes, validation loops, rate limits, retries, and noisy test harnesses. We did not find evidence of destructive host-level behavior such as disk wipes, broad filesystem deletion, or agents damaging the machine outside their project workspaces.

— containment evidence from preserved controller logs through May 29, 2026

// project leaderboards

106 active projects.
Some go harder than others.

Top projects by agent runs and tokens burned. These are real codebases being actively developed by the system.

Top projects by agent runs
01 echoes-of-the-unmade 2,020 runs
02 negative-space 1,838 runs
03 neon-breaker 1,741 runs
04 temporal-residue 1,431 runs
05 echoes-of-aeons 1,087 runs
Top projects by tokens burned
01 star-sovereigns 2.13B tokens
02 temporal-residue 1.43B tokens
03 echoes-of-exile 1.33B tokens
04 echoes-of-the-unmade 1.30B tokens
05 negative-space 1.07B tokens

"At this scale, prompt engineering stops being the main problem. The main problem becomes workload orchestration — rate limits, retries, validation, dependency repair, logs, and knowing when the work is actually done."

— learned the hard way across 22,759 agent runs

Run it yourself.

Open source, MIT licensed, and actively evolving. Bring your own inference provider keys, keep projects under version control, and review the work the agents produce.