WebinarNotes/wiki/sources/2026-07-14-skills-based-on-git.md

Git-Based Skills — the New Memory of AI Agents (Konstantin, Sber)

#source

Russian-language source. Summary in English; key terms preserved.

Source Metadata

• Date: talk (year not stated; positions "2026 (now)" and forecasts 2027)
• Raw path: raw/sources/Скиллы на базе git — новая память AI-агентов.md
• Source type: conference talk (conclusions), 53:09 — https://www.youtube.com/watch?v=a-NIeMB-Hj8
• Speaker: konstantin (Sber, GigaChat dev team, agent R&D)
• Ingestion date: 2026-07-14

Core Claims

• Skills + user data + auto-improvement = the new memory of AI agents, the architecture all agentic systems will converge on. Tools (how to call), data (what you work on), and usage history live together in one git repo; the harness decides what to load into context.
• Simplicity beats complexity: a universal agent with ~10 file-ops functions outperforms elaborate multi-agent graphs. The era of universal agents = harness.
• Agent tooling evolved tools → MCP → skills (see evolution-of-agent-tooling); skills' two-stage loading lets you keep hundreds of skills without blowing context.
• The classic skill is "von Neumann without data" — add data + usage history into the same git repo.
• agentic-loops are the next paradigm (forecast: dominant in 2027): inner (ReAct) → outer (Ralph) → meta.

Key Evidence / Details

• Harness = harness/upness metaphor: LLM = the force, tools = the harness, data = the field; the LLM "pulls the tools across the data space." Different tasks, same processing method. Standard toolset ~30–40 max (Read/Edit/Write, Bash, Grep/Glob/WebSearch, TodoWrite, subagent spawn); >100 tools = even Fable 5 gets confused. See harness.
• Human role evolution: prompt-engineer → context-engineer → harness-builder → loop-engineer (CI/CD, back-pressure, meta-loops). See context-as-scarce-resource.
• Hackathon proof: 19 of 20 top teams at BitGen built on harnesses; his own harness placed 7th (Interpress-Ex, run in an infinite loop over a weekend — read zero tasks by eye) and 3rd overall / 1st on technical metrics (Snowbase, leaked Claude Code as backend).
• tools → MCP → skills: Tools (2022–24) each consume context; MCP (Anthropic, late-2024→25) adds ready tools via protocol but many servers = dozens of tools = context blowout + network-load security risk (he doesn't hate MCP; issues largely solved). Skills (2025→) = directory + gentleman's agreement: SKILL.md + tools + (his proposal) data. Two-stage loading: short description always in system prompt; long description loaded only when the agent enters the dir. Hermes grew 56 → 100 skills and still works. Cost: MCP needs a developer; a skill needs someone who can write text (any language). Claude Code turns MCP into skills, erasing most MCP downsides. See skills-as-memory and evolution-of-agent-tooling.
• Why git: skill-first (many agents on one skill — Claude Code, Hermes via Telegram, OpenClaude anywhere) · CI/CD integration · team sharing · merge-conflict resolution now done by models (parallel actors on bare git, no transactional DB — "not for payments, but for much else, yes"). Mandatory AGENTS.md rules: git pull at start, push+commit at end, CI as back-pressure (returns a drifting agent to spec).
• His daily skills (with data): DNA/medicine (~100 GB sequencing for the whole family; Claude analyzed an undescribed mutation, found the gene, called AlphaFold, rendered the protein, concluded the fold is fine — own full genome for ~$1000 and an evening vs $3B/13 yrs historically) · recommendations (~1000 ratings) · trip planning (photo a ticket in Telegram → agent git-pulls, files the PDF, commits; "hotel not booked, 2-hr layover — don't even try") · HR/résumé scoring (self-corrected over 10–20 feedback iterations to ~10% error) · finance · team agent-R&D.
• Auto-improvement (Hermes): >5 tool calls on a task → candidate skill. Curator: prune unused at 30 days (deactivate) / 90 days (archive); consolidate pairs weekly so short-descriptions don't eat context. Memory is now skills, not RAG-extracted facts.
• Agent loops: inner = ReAct. Outer/Ralph loop (Geoffrey Huntley): while true; do claude -p "solve" || true; done — runs for days, stays in the "smart zone" (first ~⅓ of context) avoiding summarization decay. Meta loop (his own): periodically wipe all agent-created data to an unseen archive and restart, dodging Karpathy's collapse (same solution reworded) — good for research tasks.
• Auto-agent result: weak GigaChat went from 1/89 → 11/89 benchmark tasks over a weekend of self-improvement (11× gain, verified no overfit). Idea from Karpathy's "autoagent."
• Missing piece: perception / qualia (the "walk vs drive to the car wash" example); maybe world-models — open question.

Connections

• Entity: konstantin · tools: claude-code, hermes, Codex CLI, OpenClaude, Cursor, Conductor, AlphaFold
• Concepts: skills-as-memory · harness · evolution-of-agent-tooling · agentic-loops · context-as-scarce-resource
• Timeline: ai-agent-evolution
• Related sources: 2026-07-14-gap-between-ai-users-irreversible (skills = markdown folders; strong overlap) · 2026-07-14-sebastian-eugene-interview (harness, BYO-harness) · 2026-07-14-everything-we-knew-about-software-has-changed (orchestration era, markdown-as-service)

Open Questions

• What data belongs in a skill, and is there an upper limit? (No standards yet; he loads 200 GB in one skill, 100 KB in another — "haven't found the ceiling.")
• How do you build world-models / give agents perception? (Open.)
• Are the pruning windows (30/90 days) and the >5-tool-call threshold generalizable, or Hermes-specific heuristics? Status: tentative.

Change Impact on Wiki

• Created skills-as-memory, agentic-loops, evolution-of-agent-tooling; contributed heavily to harness and context-as-scarce-resource.
• Added entities konstantin and hermes; extended claude-code.
• Extended the ai-agent-evolution timeline with the tools→MCP→skills and loop generations.