root@amaltsev >:~$ cat about.md
Agent runtime security, retrieval quality, and the reliability layer underneath both, for teams that need it to work.
cat portrait.asc · 189×119
MaltsevOS 4.2 — tty1
last login: today on ttys001
Type help for the command list, or
neofetch for the machine.
↑↓ history Tab complete Ctrl+K command palette Esc clear the line
Authority, not prompting. Whether an agent may call a tool or spend a budget is a runtime decision: scoped grants, enforced budgets, durable approvals, and a record of every call.
24 notes on the seam between a model and its toolsAn answer that cannot name its passage is a guess with a bibliography. Grounded retrieval, permission-aware filtering, reranking, and a refusal when the evidence is not there.
Case studies, including permission-aware internal searchTraces, budgets, retries, and the failure states nobody designs on purpose: what an agent does when the run goes wrong, what it costs, and how anyone finds out.
5 notes on what happens when the run goes wrongHover a stage, click it for detail.
01 · BACKEND FOUNDATION FastAPI, Pydantic, AsyncIO, Pytest. The service skeleton every later stage plugs into.
02 · LLM API & TOOL CALLING Structured outputs and streaming. Every tool call is schema-checked before it is allowed to run.
03 · AGENT ORCHESTRATION LangGraph state machine. Explicit transitions, so a run can be replayed and compared step by step.
04 · API & BUSINESS INTEGRATIONS REST, webhooks, OAuth 2.0, reconciliation. The layer where the agent meets the systems of record.
05 · RAG ENGINEERING Hybrid search, reciprocal rank fusion, reranking. Retrieval quality before prompt quality.
06 · RELIABILITY & RESILIENCE Circuit breakers and idempotency keys. An upstream failure degrades the run instead of cascading.
07 · SECURITY & AUTHORIZATION Deterministic RBAC and a sandboxed tool runtime. Permissions decided in code, never by the model.
08 · EVALUATION & REGRESSION Trajectory tests over golden datasets. Every prompt or model change is measured before it ships.
09 · SQL & STATE PERSISTENCE Postgres, Redis, checkpointing. Agent state survives a crash, a deploy and a restart.
10 · OBSERVABILITY Tracing and OpenTelemetry. Every run is inspectable from first tool call to final answer.
My backend is a single async service, and every stage below is a module it loads.
asyncio.to_thread, or Celery and ARQ for parsing and vector maths. pydantic-settings loads .env and validates types at container start. psycopg2 and requests block the whole thread; I use asyncpg and httpx. asyncio.TaskGroup. httpx.AsyncClient. HTTPException and for internal 500s. pytest-asyncio fixtures with one database session per test. I only accept a model response I can validate; text I cannot parse is a bug, not an answer.
<user_context> and <documents>. "123" where I expected 123. My agent is a state graph, not a loop that I hope terminates.
interrupt_before on money moves and on outbound email. recursion_limit=10..20 stops a confused agent from burning the budget in a loop. state['key'] = val desynchronizes checkpointing and misbehaves under async. PostgresSaver. This is where my agent touches systems that charge money and send mail, so the rules here are defensive.
Idempotency-Key. PENDING, make the call, then move it to COMPLETED only on success. Idempotency-Key on every mutating call. Retrieval is a pipeline with permissions inside it, not a vector search with a prompt wrapped around it.
[Doc-12], and a post-validator checks the document really said it. acl field on every vector. Failures are a given; a run that dies on the first one is not.
tenacity or pybreaker driving the retries. The model proposes and my code decides; nothing runs that the backend has not authorized.
min(limit, 50), so a single call cannot dump the database. Every prompt change is a measurement, and until it is one it is a guess.
promptfoo or pytest on every pull request and every prompt change. A conversation that vanishes when a process restarts is a bug, so state lives in Postgres and only the hot path lives in Redis.
m=16, ef_construction=64, and the right operator class for the distance. When a run goes wrong I need to know which request, which node and which call produced it.
Four channels, all of them read. Pick the one that suits the question.
Where to start: a short question goes to Telegram, a brief goes to email.
No match.