Part of pureWRIGHT · Engineering Record

pureBOT

The part that does the work.

pureBOT takes a task, runs it on the model the operator chose or the local default, calls the tools it needs and keeps an audit record. Where the operator has reserved a decision, it refuses and says what it needs, and the operator decides. It works inside PureTensor's own operations today. Below is its engineering record, including how it grew out of a small personal agent.

DOCUMENT REV 2026-10-08STATUS: OPERATIONAL
The Film · 3 min 41 s

How pureBOT does the work

Four ways in, one model per request, pureBOT's own tools behind one executor, refusals that leave the decision with the operator, observers in the background, and an audit record. Every frame was generated by code and rendered on our own GPUs; the figures on screen come from the running code and its audit log, checked on 8 October 2026.

Models
The default model is open-weight and runs on GPUs PureTensor owns; which model sits in that seat changes from time to time. When it is rate-limited, down or too slow, a request fails over to a hosted frontier model, such as Claude on Amazon Bedrock. The operator can also ask for a larger Claude model, or for the Claude Code or Codex agent.
Boundary
A request that fails over, or that the operator sends to a hosted model, leaves our hardware with its context. The Claude Code and Codex agents run their own tools: pureBOT records the model call, not each step they take.
Record
The counts come from pureBOT's own audit log for the 30 days to 8 October 2026. The tool-call count includes calls the policy refused.
Section 01 · Origin

The agent that outgrew itself

2024 · First commit. A reaction to bloated agent frameworks.

The monolith began as a reaction. The first wave of open-source agent frameworks had proven the pattern (markdown identity files, tool use, multi-channel ingress) but brought mass dependency, untested security boundaries, and an abstraction tax that made debugging impossible.

Ours stripped the idea to its essentials: a minimal agentic harness around Claude, with a three-tier memory system, a multi-engine inference router, and a policy-driven security framework. The goal was a personal AI agent you could actually trust on production infrastructure.

Identity
Three-tier markdown memory
SOUL.md, USER.md, MEMORY.md. Versioned, diffable, grep-able.
Inference
Multi-engine routing
Claude as the default cortex; local and cloud backends behind it.
Safety
Policy-driven security
An explicit allow-list, redaction pipeline, and audit trail around every tool call.
Section 02 · Absorption

From agent to nervous system

2025–2026 · The chatbot stopped being the point.

The inflection point came when the monolith’s operational modules started doing more useful work than the chatbot interface they were built to serve.

One by one, the modules were absorbed into the infrastructure. The chatbot was the last thing standing, and it was the least useful part.

01
Observer system
11 scheduled observers
Operational awareness on schedules of its own: infrastructure health, rail telemetry and repository review, reported to the operator.
02
Heuristic dispatcher
10 pre-LLM data APIs
Sub-second fast path for infrastructure queries. Structured answers without spinning up a model.
03
Multi-engine router
7 backends, automatic failover
One inference contract for every pureBOT lane: local GPU, cloud API and CLI agents behind it.
04
Security framework
6 policy modules
Baseline security posture: policy engine, audit trail, credential redaction, SSRF protection, filesystem ACLs, inference guards.
Section 03 · pureBOT

The nervous system

Naming the operational reality that was already there.

pureBOT is what remains after the agent identity is removed and the operational reality is named. It coordinates autonomous infrastructure modules across PureTensor’s compute fleet.

The biology analogy is deliberate. pureBOT is the autonomic nervous system of the cluster. It does not think. It coordinates.

Model selection and failover
Default · Local
Open-weight model on our own GPUs
Handles every request unless a chat or a job asks for another model. On-fleet, no per-token cost.
Fallback · Cloud
Claude via Amazon Bedrock
Takes over automatically when the local model fails. Larger Claude models, and agent CLIs, on request.
No model · Data cards
Slash commands
Status, markets, rail and node-health cards answer from structured data. No model in the path when one is not needed.
Scheduled observers

Background jobs watch the estate and external signals on their own schedules and report to the operator. Each is a self-contained module with its own cadence.

Built-in tools

Shell, file and web tools run through one executor, which applies the security policy to every call.

Task scheduling

A scheduler runs tasks and reminders. A heartbeat gathers estate status on a schedule and logs what it finds.

Audit and limits

Every tool and model call is logged. Risky commands, such as deletions and service stops, are blocked by rule, and a daily ceiling caps paid API spend.

Section 04 · Architecture

How the pieces fit

Messages from Telegram, Discord, email and the terminal go straight to the core. Background observers run on their own schedules and report to the operator; the briefing they keep feeds the core's context. The core uses one model per request, the local model by default, and falls back to Claude on Amazon Bedrock if that fails. It draws on built-in tools and markdown memory, and every action is audited.

Security boundaryPolicy, audit, redaction, risky-action blocking.
Ingress
  • Telegram
  • Discord
  • Email
  • CLI
ObserversScheduled jobs that brief it
CorepureBOT
SchedulerTasks and reminders
One model per request
One model per request
  • DefaultLocalopen-weightown GPUs
  • FallbackCloudclaudebedrock
  • On requestAgent CLIsclaude codecodexgemini
draws on
  • Toolsshellfilesweb
  • MemorymarkdownSQLite
  • Auditevery call loggedmetrics
Scheduled heartbeat, logged
Section 05 · What Survived

Lineage

What carried over, what got renamed, what got cut.

Most of the monolith is still running. It just stopped pretending to be one thing and started being many. The table below is the continuity map: who became who.

Monolith observer system
Always-on monitors
pureBOT scheduled observers
Fewer, headless, on their own schedules. They report to the operator.
kept
Monolith heuristic dispatcher
10 pre-LLM data APIs
pureBOT fast-path query layer
Same dispatch table. Slash commands answer from data, with no model.
kept
Monolith multi-engine router
9 backends, automatic failover
pureBOT model selection and failover
One model per request: local by default, Claude on Bedrock if it fails.
renamed
Monolith security framework
Policy, audit, redaction, ACLs, guards
pureBOT policy engine
Same modules, plus risky-action blocking and a daily spend ceiling.
kept
Monolith memory (SOUL/USER/MEMORY.md)
Three-tier markdown identity
pureBOT durable knowledge layer
Markdown files preserved, with SQLite for sessions, history and the audit log.
renamed
Monolith multi-channel ingress
Telegram, Discord, WhatsApp, email, CLI
Channel adapters (HAL identity)
Telegram, Discord, email and the terminal remain live. No longer the centre of the system.
renamed
Monolith chatbot interface
Conversational front door
—
Discontinued. The interface was the least useful part.
discontinued

A project that changed shape, not a product launch.

From the Lab

The research behind pureBOT

End of Document

Built, running, and yours.

pureBOT is the part of pureWRIGHT that does the work, alongside pureMIND, which remembers everything, and pureMEND, which repairs what breaks. It runs PureTensor's own operations today and is deployed for enterprise inside your boundary. There is no pricing page; an engagement begins with a conversation with the engineers who run it.