Vladimir Koptsev

what it is

Before delegation, each task gets a class (A mechanical, B heavy compute, C closed or open, D judgment, voice, money, outbound) and a model. Fable plans and accepts key work, Opus reviews and runs open tasks, Sonnet executes, Haiku handles small mechanical jobs. A script assigns the class and model from a rules table without calling any model, so routing costs zero tokens.

how it works

The routing script (task-router.py) has the commands route, outcome, solo, consult, calibrate and status. It learns from verdicts: two failures in a row move a task type one step up the ladder. Moving down is allowed only for classes A and B after five clean accepts, and a rollback freezes the rule for 30 days. Route log: 1965 entries.

in a company

Rules pick the model for each task so the token bill stays under control. The expensive model sets the task and accepts the result, cheaper ones do the work.

human check

I or an Opus reviewer accept each subagent result; the verdict goes into the log and changes future routes.

result

Week to 12 October: 161 delegations (Sonnet 81, Opus 75, main session 5), 130 accepted, 2 sent back for rework, 1 escalation, 3 failures, 23 consultations with the senior model. Orchestration share 96%.

ladder: haiku -> sonnet -> opus -> fable; classes: A=sonnet, B=sonnet, C1=sonnet, C2=opus, D=session
learned rules: 4
last 7 days: 161 delegations (opus:75, session:5, sonnet:81); verdicts: {'ok': 130, 'retask': 2, 'escalated': 1, 'failed': 3}; rule changes: 1
orchestration share: 96% (5 solo); consultations with the senior model: 23

what it is

No email, message or job application leaves my system without a ticket. A script issues the ticket when the conditions line up: the recipient is known, the text has no money and no unverified third-party facts, and I have said "ok".

how it works

Three ticket types. Red covers one recipient for 30 minutes and needs my "ok" quoted word for word. Batch lasts 6 hours and works only for a fixed address list. Green lasts 60 minutes and needs one of three grounds: an approved template, a reply in a live thread or a short service reply. The risk light (risk-gate.py) defaults to red and decides by rules, with no model involved. It learns from verdicts and incidents, and an incident freezes the ground it slipped through. Claims map: the script refuses a red ticket until every third-party name and number in the email has a line saying where it came from and who confirmed it. Before drafting, the agent rereads the recipient file: what we already wrote to them, what we promised, what we turned down. Follow-up corrections are banned: once an email is sent, nobody amends or retracts it.

in a company

Sales emails to clients: the agent writes, a script checks the recipient and the figures, an employee presses "send".

human check

Any money marker in the text stops the send until a senior model has reviewed it and I have given a word-for-word "ok".

result

Risk light selftest 15 of 15; 867 risk-light decisions in the log; 235 verified claims (names, sums, dates) in recipient files. A send counts as done only when the receiving side shows it (the Sent folder, the channel itself).

[PASS] service reply, green: exit 0 (expected 0)
[PASS] plain amount of money: exit 2 (expected 2)
[PASS] masked amount (NBSP+ZW): exit 2 (expected 2)
[PASS] red marker "EUR sign": exit 2 (expected 2)
[PASS] red domain from the government list: exit 2 (expected 2)
[PASS] red address in cc: exit 2 (expected 2)
[PASS] service reply with two recipients: exit 2 (expected 2)
[PASS] thread with no history: exit 2 (expected 2)
[PASS] thread with clean history: exit 0 (expected 0)
[PASS] thread with red history (any_red): exit 2 (expected 2)
[PASS] unregistered template: exit 2 (expected 2)
[PASS] telegram channel: exit 2 (expected 2)
[PASS] long service reply: exit 2 (expected 2)
[PASS] attachment in a service reply: exit 2 (expected 2)
[PASS] daily limit: 20 green, the next one exits 2 (expected 2)

SELFTEST PASS: 15 scenarios, live state untouched

what it is

Checks that the environment runs before and after agent actions. A hook fires before a tool call and blocks it when a condition fails.

how it works

Outbound. The hook blocks an email without a fresh ticket. It also blocks publishing, editing or deleting a post on social networks and Telegram without a ticket. outbound-guard, publish-guard

Token budget. A subagent gets a warning at its 60th call and loses its tools at the 100th. Subagents go to the background only when told to. Screenshots shrink to 1568 px, and PDFs are read as text, not as images. subagent-step-limit, agent-foreground-default, shrink-image, pdf-to-md

Housekeeping. The agent does not create files "for the owner to read"; content goes to the chat. Every turn gets an exact timestamp. A changed skill triggers a registration reminder, and the knowledge base syncs after each session. reader-guard, now-stamp, skill registration, knowledge base sync

in a company

Agent restrictions live in the environment settings and run before every action: a sales agent cannot send an email without an issued permission.

human check

The hook returns a refusal that says what to get (a ticket, a quoted "ok"), and the turn comes back to me.

result

Without a ticket, the hook blocks the call to the send tool. Lesson from the 19 September 2026 audit: one invalid line in settings.json silently disabled the whole file along with its hooks. Since then a script (claude doctor) checks the hooks after every config edit.

PreToolUse ['pdf-to-md-guard.sh', 'shrink-image-hook.sh', 'publish-guard.py', 'outbound-guard.py', 'reader-guard.py', 'subagent-step-limit.sh', 'agent-foreground-default.py']
PostToolUse: reminder to register a changed skill
Stop ['claude-obsidian-sync.sh']
UserPromptSubmit ['now-stamp.sh']

Several models read every important decision and every outbound text before I accept it. The internal council follows the DMAD scheme: five advisor subagents with different roles, a devil's advocate and a chair, and confidential material can go there. External panels take non-confidential work: six model families on NVIDIA NIM and six models on OpenRouter, with Codex joining as a second agent through an MCP bridge I wrote. A monthly smoke test checks that the models still respond. I write the synthesis and make the call. A council of 12 agents that wrote 15 rules into a skill took 15 minutes (measured), and a copy of the panel runs on the server for Hermes.

what it is

Open-source tool that lets Claude Code or Codex work inside an editing program as an assistant editor. I built it for my production company and published the code.

how it works

Node MCP server with ten pr_* tools: status, project, sequence, timeline, selection, playhead, markers, export to Adobe Media Encoder, arbitrary ExtendScript. Python core of 25 modules: silence and best-moment search, audio mixing with ducking, subtitles and transcript checks, color gate and technical file QC, edit variants and revisions to a hand-built cut, XML handoff. Adapters for Premiere, DaVinci Resolve (Studio Python API) and FCPXML. Bundle for Claude Desktop and server.json for the MCP registry.

in a company

An editor hands routine work inside their own program to the agent: logging, finding pauses, checking a file before delivery. Meaning and rhythm stay with the editor.

human check

The director decides the meaning and rhythm of the cut. Every error from a real job becomes a code check with a test.

result

28 test suites, 707 checks, ALL GREEN on 12 October 2026; tested on real footage: 108 source files, 4.4 GB.

xml_handoff: 11 passed, 0 failed

ALL GREEN

A round-the-clock agent that keeps working when my laptop is off: it answers people in support bots, keeps the family books and sends morning digests. It runs as a separate user without root, with one narrow wrapper as its only way into sudo. Approval mode is set to smart, the no-confirmation mode is banned, and mail keys are not in its environment. Nine isolated profiles split the work, all on free models through NVIDIA NIM (Nemotron 3 Ultra, with DeepSeek and Gemini Flash as fallbacks). A bridge script on my Mac (hermes-task.sh) hands it tasks without spending main-session tokens, and it passes hard questions to me in a staff chat. The server costs about 7 euros a month at list price. One profile, Kassa, is a family bookkeeping agent: two models read every receipt independently.

what it is

One panel runs my Telegram channel, VK, Instagram and LinkedIn. The panel has no "publish" button: a script publishes once there is a ticket and my "ok".

how it works

FastAPI and SQLite panel on the server, with an in-process scheduler, a calendar and a text version for each platform. Only I can approve a post, by a tap in the Telegram bot, a word in the agent chat or the "Approve" button on the post page; any edit to the text, image or platform list removes the approval. Before drafting, the agent runs a TF-IDF repeat check against every channel post and the topic map. Before publishing, a script checks the text (text-gate.py: personal stoplist, platform limits, markup) and issues a 30-minute ticket. After publishing, a check script (tg-verify.py) looks at the real rendered post, because an API "ok" does not count as proof. The feed guard (tg-scan.py) rereads the whole feed after 2 and 15 minutes, and a duplicate means stop and report. A long post with an image goes out from my own account through Telethon (a user-account Telegram client), only with a ticket and my "ok". A nightly job saves view counts and a morning card sums up the channel.

in a company

A content team where one person holds the button: agents prepare posts for each platform, and only the manager releases them.

human check

The agent sends me the full post text in chat. I read all of it and write "ok" for specific platforms. Then the script issues a ticket and the pipeline publishes.

result

621 posts of history imported (146 Telegram, 475 VK); 212 posts in the repeat-check base. I dropped the stylometric detector after an audit showed 18-20% false positives on my own texts and replaced it with the repeat check and topic map.

what it is

The agent writes in Russian in my voice, and I measure acceptance by draft survival: the share of the draft that stays after my edit. It is 84% over the last ten texts, against a target of 90%.

how it works

Corpus of 25 confirmed texts of mine (about 6400 words) with a provenance map that separates my own voice from character speech and quotes. I log each edit as a dated pair, "agent's phrase, my phrase", and error classes come out of the pairs: 6 classes of things I cut, 12 classes of things I fix. A class needs at least two examples, and the agent writes down what an edit means only after a short interview with me. Stoplist: 15 hard rules and 12 warnings, all from my direct bans. Before showing a draft, the agent runs a "stranger test": a fresh context checks every verb-plus-object pair with three questions.

in a company

Letters and posts in a manager's name: the agent writes in their voice, and the share of edits shows how close it gets.

human check

I edit, and the system learns from the pairs; a whole draft never becomes a reference text.

result

47 dated pairs in the log. Draft survival: target 90%, now 84%, log open.

average over the last 10 drafts: 84% (target >= 90%)