Releases, method notes, and how we think about digital employees that grow.
6 min read
We read the sub agent implementations in three open source agent projects and compared their choices on context isolation, permission derivation, recursion guards, concurrency, timeouts and return contracts. All three are equally narrow, and none has a timeout or a retry. We close with what this means for digital employees delegating to each other.
6 min read
We read the Manus team's public material on context engineering, from the 2025 blog post to the later workshop and interviews, extracted the checkable principles and thresholds, and compared them with how createrole compresses context on the server today. Five principles are aligned. One is not.
3 min read
RSI went from thought experiment to engineering mainstream, yet almost everything that actually runs is bounded self-refinement. createrole stands firmly on that side. Model weights stay frozen. What changes is memory, notes, skills and the exam system, with a human on the loop and every step reversible.
3 min read
Using a data analyst employee as the example, we walk the loop as it exists in the product. Generate a question bank by profession, score a baseline with a judge model, coach a candidate on a snapshot, and let a human decide before it enters version history.
5 min read
A digital employee cannot grow on user feedback alone. Night school generates a syllabus and a tiered, anchored question bank per profession, separates the examiner from the student, splits practice and hidden pools, and turns eval scores into the external anchor for coaching. This post explains the design choices behind it.
9 min read
A survey of RSI research and practice from 2025 to 2026, covering definitions, representative work, product-grade growth without training weights, supervision and evaluation, and the failure modes the literature keeps repeating. It closes with how createrole applies these findings to the growth loop of a digital employee.
10 min read
Read side by side, the 2025 to 2026 official agent prompting guides from Anthropic and OpenAI point the same way. Delete instructions instead of adding them, restrain the model instead of pushing it, and write goal contracts instead of step lists. This post collects the most important shared positions and the most actionable rules.
14 min read
A source-level comparison of how three open-source agents, Codex, Pi and OpenCode, summarize old history, how much recent verbatim text they keep, and how they recover from overflow. The takeaway is that compression is really the construction of a recoverable state checkpoint, not a summary of a chat log.
9 min read
We read the source of three open-source agent memory frameworks and compared how they write, resolve conflicts, forget and retrieve. All three have converged on time anchoring, lexical channels and provenance. Forgetting is still the least mature part.
