AI swarm tooling

An AI lab building next-generation swarm tooling for planning, building, and verifying software.

MiniFlow

A Python 3.13 workflow engine built from the standard library: validation, deterministic concurrent scheduling, retries, timeouts, cancellation, checkpoint and resume, events, and a CLI. Solo and swarm figures are medians from three predeclared unattended paired runs. The adaptive figure is the later resource-aware attempt 07, a single formally valid diagnostic run against the same frozen fixture and hidden acceptance matrix.

Mode	Accuracy	Reliability	Active AI time	Cost
Solo	44.4%	33%	10m 30s	$0.045 (best cost)
Swarm	97.0%	67%	13m 30s	$0.297
Adaptive	100.0% (best accuracy)	100% (1/1) (best reliability)	6m 41s (best active ai time)	$1.330

Adaptive attempt 07 achieved full correctness in 6m 41s for $1.330, making it the fastest and most accurate displayed result, but it is one diagnostic observation rather than a repeated-run estimate. In the earlier paired matrix, swarm was more accurate and reliable than solo but cost 6.6 times solo's median spend; both modes were highly variable across runs.

RedisClone

A single-node Redis-compatible server: RESP protocol parsing, string and hash commands, key expiry, persistence snapshots, and a concurrent client loop. Figures are from the pilot pair, extended with demo placeholder data pending the full three-pair matrix.

Mode	Accuracy	Reliability	Active AI time	Cost
Solo	96.5% (best accuracy)	100% (best reliability)	10m 48s (best active ai time)	$2.07 (best cost)
Swarm	95.5%	100% (best reliability)	26m 54s	$8.61
Adaptive	96.2%	100% (best reliability)	12m 06s	$2.84

Solo matched the swarm's accuracy on this fixture at roughly a quarter of the cost and well under half the time, and the adaptive router selected solo accordingly. Placeholder conclusion pending the full matrix.

SQLite-in-Rust

A minimal SQLite-compatible storage engine in Rust: file-backed B-tree pages, a bytecode query planner for a SQL subset, transactions with rollback journaling, and a REPL. Figures are demo placeholder data pending the first predeclared matrix.

Mode	Accuracy	Reliability	Active AI time	Cost
Solo	62.1%	33%	24m 12s (best active ai time)	$1.12 (best cost)
Swarm	78.3% (best accuracy)	67% (best reliability)	31m 45s	$4.95
Adaptive	76.9%	67% (best reliability)	27m 30s	$3.10

The swarm held a clear accuracy and reliability lead on this harder fixture, and the adaptive router captured most of that gain at well under the full swarm cost. Placeholder conclusion pending real runs.

Capd-from-docs

A capability daemon implemented solely from its written specification: policy parsing, capability grants and revocation, an audit log, and a Unix-socket API, with no reference implementation available. Figures are demo placeholder data pending the first predeclared matrix.

Mode	Accuracy	Reliability	Active AI time	Cost
Solo	88.9%	67%	8m 15s (best active ai time)	$0.31 (best cost)
Swarm	91.2% (best accuracy)	67%	14m 40s	$1.48
Adaptive	90.1%	100% (best reliability)	9m 05s	$0.52

Accuracy was close across all three modes on this specification-driven fixture; the adaptive router reached the swarm's reliability at near-solo time and cost. Placeholder conclusion pending real runs.

Journal

Notes and learnings from our work, published roughly weekly.

07 August 2026

Email hello@jxd.dev to find out more.