Miadi Review

Staged Agent Inquiry and Skill Evolution via Hermes Gateway

Oct 15, 2026 · model: manual-import · channel: JGWill

Watch source

Staged Agent Inquiry and Skill Evolution via Hermes Gateway

Source: screenwalk-kherix-2610141715-a-room-that-can-hold-the-work-261014 — JGWill https://youtu.be/EnoUOUxRqRU

Summary

An operator directs a Hermes agent through Telegram to set up a staged inquiry-to-script workflow inside the Miadi orchestration framework. The operator steers the agent to define the T5 production context exploration pattern, in which a workspace holds unfinished research and prompts and creates no official episode. The agent inspects repositories, revises five skill files, runs an ad-hoc Python script of 12 assertions, commits the updates, and hands the results to Scrivener documentation and ElevenLabs speech synthesis.

Reading notes. The figures, timecodes and order of events in this review are as summarized from the video, not checked. The video runs 41:38 (YouTube page metadata, 2498 seconds), so no timecode below passes its end. The generated text also disagrees with itself on order: the ad-hoc test sits at 22:30-24:30 under Aspects and at 39:18-40:05 under Results, and the Scrivener and ElevenLabs hand-off (34:20-38:15) comes before the second. Either two test runs or a mis-timed summary would fit; only the video settles it.

Academic Fields

Human–AI Interaction (Human–Computer Interaction)

Studies how people understand, direct and correct AI systems. The video touches it as an aspect: the operator steers the agent mid-run from a phone through /steer notes. Amershi et al. (2019) list "support efficient correction", but tested it by heuristic evaluation of existing products, not agent steering.

Human Oversight in Software Process (Software Engineering)

Studies where people must decide inside automated work. The video touches it as a result: the agent drafts, a script checks the edits, and a person keeps the choice to produce an episode. Parasuraman et al. (2000) grade automation by stage. Bainbridge (1983) limits the claim: a human step shows presence, not vigilance.

Provenance and Research Data Management (Information Science)

Studies keeping the origin, custody and revision of data recoverable. The video touches it as an aspect: the staging folder keeps source, interpretation and correction apart and records no episode number or receipt. PROV-O (W3C 2013) models entities, activities and agents, but a provenance record shows what was recorded, not that it is true.

Narratology and Transmedia Storytelling (Literary and Media Studies)

Studies how stories are built across media. The video touches it as an aspect: a script drafted before any episode, then spoken through ElevenLabs. Genetic criticism (Deppman et al. 2004) reads drafts as part of the work. Jenkins (2007) wants each medium to add something unique, so reading text aloud is adaptation, not extension.

Indigenous Research Methodology (Research Is Ceremony, Wilson 2008)

Studies research as accountability to relations: who takes part, who interprets, who benefits. The video touches it at one point: the decision to produce an episode is left to a person, not the agent. Wilson (2008) places that accountability with people in relation, so a staging rule can respect it but cannot discharge it.

Language Agents and Skill Libraries (Machine Learning)

Studies agents that keep reusable procedures and improve them from feedback. The video touches it as a result: five skill files revised, then checked by a script. Wang et al. (2023) verified skills against game feedback in Minecraft. Sumers et al. (2023) rate writing to procedural memory significantly riskier than writing to other memories.

Aspects

Hermes Gateway Mobile Orchestration (00:30-01:30)

Operating through Telegram with the Hermes Gateway running on OpenAI Codex, the operator monitors live session context, checks token consumption, and manages platform bindings from a mobile device (as summarized from the video, not checked). Hermes's own documentation lists Telegram as a gateway platform, OpenAI Codex as a provider option signed in with ChatGPT credentials, and /usage and /status for token use and session information. Those are the vendor's statements, not the video's.

Interactive Execution Steering (05:40-07:05)

Using /steer commands, the operator redirects the agent's task in flight, corrects speech-to-text recognition errors such as 'Kima' (noise that names no research subject), and points exploratory research at designated repository paths. Hermes documents /steer as a note appended to the last tool result once the current tool completes, with no interrupt and no new user turn. A steer is therefore text the model reads next. It is not a limit the runtime enforces.

Decoupling Staged Context from Episode Creation (15:10-17:40)

The workflow separates exploratory ideation from episode birth. Staging files go inside an underscore-prefixed directory and hold research without an episode manifest or an episode number. The staging folder this video concerns states the same boundary in writing: no episode number, manifest, registration receipt or wheel node is authorized, and the choice to produce an episode stays with a person. The boundary is a rule the agent follows. Nothing in the summary shows the system refusing a mint.

Targeted Ad-Hoc Verification Scripts (22:30-24:30)

The agent writes, runs and deletes an ad-hoc Python script that checks the updated skills for boundaries, word limits and repository commit scopes (as summarized from the video, not checked). A script of this kind checks form and scope. The staging folder's own verification record lists ten named checks: frontmatter on the five skills, word bounds, whitespace in the scoped diffs, registration of the shared skill, and no episode manifest or receipt. It also states that no browser rendering or full repository test suite ran. The video's 12 assertions and the record's ten checks count different things, and neither confirms the other. None of them runs a revised skill in use.

Multimodal Curation and Speech Synthesis (34:20-38:15)

The generated logs and staging synopses are transferred to Scrivener and Apple Notes, then spoken with ElevenLabs speech synthesis, which gives the exploration a form a person can read and listen to (as summarized from the video, not checked). Hearing a log read aloud supports a human review. It does not show that the review took place.

End-results and Core Aspect

Core Aspect — Staged Production-Context Exploration (15:10-22:30)

The workflow separates experimental generative work from episode finalization. The orchestration kit gains a staging practice, written into skills, so that agents hold raw notes, prompts and screenwalks without registering an episode. The separation depends on agents following the written rule, and incomplete creative work stays preserved until a person chooses.

Results

  • Five Skill Bodies Revised (15:10-18:30): As summarized from the video, not checked, the Hermes agent modified and committed five skill definitions across multiple repositories, adjusting routing guidelines, MCP CLI procedures and context-explorer roles to formalize pre-episode incubation. The staging folder's change map counts five skill bodies, one new and four revised. They carry four skill names, because chronicle-episode appears twice (the shared kit copy and the Kherix adaptation), and they sit in two places, the shared orchestration kit and the default Hermes profile. Commit and push receipts are held in the issue handoff and were not checked here.
  • Ad-Hoc Test Gate (39:18-40:05): As summarized from the video, not checked, a temporary Python script validated the changed skills, formatting limits and human-ownership metadata, passed 12 targeted assertions, and removed its own harness. A pass of 12 assertions shows that those 12 held in that run. It does not show that the skills work in use.
  • Cross-Platform Multimodal Synthesis (34:20-38:15): As summarized from the video, not checked, the verification log and staging synopsis were handed to Scrivener for documentation and synthesized into speech with ElevenLabs. This shows one hand-off path from agent output to a person's tools. Auditability would need a record of who reviewed what, which the summary does not give.

Sources

Tags: fetched means the page or paper was opened in this session. search-verified means a web search confirmed the bibliographic details and the source was not opened. from memory means neither.

  • Amershi, S., Weld, D., Vorvoreanu, M., Fourney, A., Nushi, B., Collisson, P., Suh, J., Iqbal, S., Bennett, P. N., Inkpen, K., Teevan, J., Kikin-Gil, R., & Horvitz, E. (2019). Guidelines for Human-AI Interaction. CHI 2019. https://doi.org/10.1145/3290605.3300233 (fetched: full text; G9 wording and the 49-practitioner, 20-product modified heuristic evaluation checked)
  • Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775-779. https://doi.org/10.1016/0005-1098(83)90046-8 (search-verified: bibliographic details; the claim about operator skill and monitoring is from memory and secondary summaries)
  • Deppman, J., Ferrer, D., & Groden, M. (Eds.). (2004). Genetic Criticism: Texts and Avant-textes. University of Pennsylvania Press. ISBN 9780812237771. (search-verified: catalogue records; the book was not opened)
  • Jenkins, H. (2007, March). Transmedia Storytelling 101. Confessions of an Aca-Fan. https://henryjenkins.org/blog/2007/03/transmedia_storytelling_101.html (fetched: the sentence "Ideally, each medium makes it own unique contribution to the unfolding of the story" quoted from the page, spelling as published)
  • Lebo, T., Sahoo, S., & McGuinness, D. (Eds.). (2013). PROV-O: The PROV Ontology. W3C Recommendation, 30 April 2013. https://www.w3.org/TR/prov-o/ (fetched: abstract and definitions of entity, activity and agent)
  • Parasuraman, R., Sheridan, T. B., & Wickens, C. D. (2000). A model for types and levels of human interaction with automation. IEEE Transactions on Systems, Man, and Cybernetics, Part A, 30(3), 286-297. https://doi.org/10.1109/3468.844354 (search-verified: DOI from a BibSonomy record, the IEEE page was not opened; the staged account of automation by stage is from memory)
  • Sumers, T. R., Yao, S., Narasimhan, K., & Griffiths, T. L. (2023). Cognitive Architectures for Language Agents. arXiv:2309.02427 (TMLR, 2024). https://arxiv.org/abs/2309.02427 (fetched: abstract; the procedural memory passage in section 4.1 read in the PDF text)
  • Wang, G., Xie, Y., Jiang, Y., Mandlekar, A., Xiao, C., Zhu, Y., Fan, L., & Anandkumar, A. (2023). Voyager: An Open-Ended Embodied Agent with Large Language Models. arXiv:2305.16291. https://arxiv.org/abs/2305.16291 (fetched: abstract; environment feedback, execution errors and self-verification, evaluated in Minecraft)
  • Wilson, S. (2008). Research Is Ceremony: Indigenous Research Methods. Fernwood Publishing. ISBN 9781552662816. (search-verified: catalogue records and secondary reviews; the book was not opened, so the account of relational accountability rests on those and on memory)
  • Nous Research. Hermes Agent documentation: slash commands (/steer, /usage, /status), https://hermes-agent.nousresearch.com/docs/reference/slash-commands ; messaging gateway (Telegram), https://hermes-agent.nousresearch.com/docs/user-guide/messaging ; providers (OpenAI Codex), https://hermes-agent.nousresearch.com/docs/integrations/providers (fetched 2026-10-14; vendor statements)
  • YouTube watch page for the video (length 2498 seconds). https://www.youtube.com/watch?v=EnoUOUxRqRU (fetched 2026-10-14)
  • Staging folder a-room-that-can-hold-the-work-261014 under the Chronicle's _staging_for_new_episodes/: AGENTS.md, skill-evolution.md, .mia/verification.json, .mia/academics.md (read 2026-10-14; internal records of the work, not independent confirmation of what the video shows)