Curated research from Kiloloop — runtime comparisons, operational reviews, practical guides, and protocol patterns for multi-agent coordination.
Research published from our internal pipeline. Each piece is reviewed and sanitized before landing here.
| Directory | Contents |
|---|---|
runtime-comparison/ |
Runtime capability comparisons and benchmarks |
ops-reviews/ |
Operational reviews of our own agent fleet, with the scripts behind the numbers |
practical-guides/ |
How-to guides for agent coordination workflows, written from our own runs |
| Date | Piece | Where |
|---|---|---|
| 2026-11-12 | Always-on agent interop matrix — what OpenAI dot, xAI Grok Bot and Claude Duet can do and how work gets in and out of each: channels, triggers, connectors, memory, shared storage, approvals and auth, self-reported by the agents with an evidence tag on every claim | runtime-comparison/ |
| 2026-11-01 | Claude Code in the cloud — a 101 from our own runs — which job shapes cloud sessions and routines fit and which broke, the three pieces (environment, session, routine), a setup order that worked, nine things that surprised us, and what the runs cost, from a week of test runs on the research preview | practical-guides/ |
| 2026-10-01 | Validated dispatches, defined (second edition) — the same four-test definition applied to the record through 2026-09-30: both scripts re-run, the board reader corrected and the first edition's dates re-read beside it, with the first edition kept as published; new aggregates and a redacted sample ship alongside | ops-reviews/ |
| 2026-09-10 | Validated dispatches, defined: what the number on our homepage counts, and what it leaves out — the definition behind the dispatch count: four tests, the exclusions, the count over time with its correction to 557 dispatched legs, what the count is made of, and failure and human-intervention criteria from both records; aggregates, a redacted sample and CITATION.cff ship alongside, and the numbers come from estimate_stats.py and hil_stats.py |
ops-reviews/ |
| 2026-09-07 | Estimate accuracy, measured: seven months of dispatch estimates against their actuals — actual-to-estimate ratios from 606 dispatch rows by period, month, estimate size and task class, the declared-envelope headroom on review-loop legs, verdicts and review status; the computation behind the numbers is estimate_stats.py |
ops-reviews/ |
| 2026-09-05 | Human in the loop, measured: four months of autonomy-gate records — pause rates, human outcomes and latency, checkpoint breaches and declared-vs-actual fidelity from 602 audit records; reproducible with hil_stats.py |
ops-reviews/ |
| refreshed 2026-11 | Runtime capability matrix — what each coding-agent runtime can and cannot do, self-reported and cross-verified by the agents | runtime-comparison/ |
| refreshed 2026-09 | Prompt caching patterns — Claude and Codex caching defaults, billing mechanics, and measurement | runtime-comparison/ |
Code: Apache License 2.0 Content: CC BY 4.0