Custom Agents
A focused assistant or operator built around one real job, with tools, limits, escalation, and ownership made explicit.
01 CUSTOM AI SYSTEMS / RECURSIVEINTELL
AI systems built
around your business.
I design and build custom agents, workflow automation, business knowledge systems, and tool integrations around the work your business already does. Local-first options, human approvals, and traceable execution stay explicit.
Local-first optionsHuman approvalsTraceable execution
02 / WHAT I BUILD
Useful systems.
Clear boundaries.
Start with a business problem, not a generic bot. Every system names its source data, tools, decision points, human owner, and failure behavior.
Workflow Automation
Repeated follow-up, reporting, document, intake, or coordination work mapped before any part of it is automated.
Business Knowledge
Source-aware answers across the documents, policies, notes, and operational context your team is allowed to use.
Tool + Data Integrations
Carefully bounded connections between AI workflows and the software, files, APIs, or local systems you already depend on.
03 / FOCUSED CONSULTING
Judgment first.
Then the build.
Not every useful engagement needs a new product. I can map the workflow, review an agent architecture, pressure-test a local-first design, or define the evidence gate your team needs before committing to implementation.
See consulting engagements →Agent runtime architecture
Graph design, model and tool boundaries, memory routes, failure states, receipts, and operator control.
Hermes integration
Skills, MCP servers, local memory, Rust-backed services, multi-agent workflows, and bounded performance work.
Local-first AI
Deployment topology, data residency, provider boundaries, offline behavior, recovery, and explicit degraded modes.
Reliability + proof
Hostile workflow review, replay cases, claim/evidence boundaries, acceptance tests, and release gates.
04 / SEE THE SHAPE
From input to action.
Nothing hidden.
Choose a common workflow to see where evidence, approval, action, and failure should remain visible.
Website form or shared inbox
Classify the request and draft the next response from approved business facts
A person reviews consequential or low-confidence replies
Send through the existing email or CRM tool after approval
Record the input, approval, action, and explicit failure state
Example workflow, not a customer deployment claim.The example does not promise autonomous sales decisions or customer results.
05 / HOW WE WORK
Map. Build.
Verify. Operate.
Consulting and implementation use the same discipline: current evidence first, one canonical owner per concept, mapped acceptance gates, and a written remaining delta.
- 01
Map
Name the repeated work, current owner, inputs, exceptions, and decision points.
- 02
Build
Implement the smallest useful system without hiding provider, data, or authority boundaries.
- 03
Verify
Test the agreed path, failures, approvals, and handoff against explicit acceptance conditions.
- 04
Operate
Document ownership and support the system only to the level the engagement actually requires.
06 / CONTROL STAYS EXPLICIT
Local when useful.
Connected when required.
Never blurred.
Local-first options can keep durable knowledge and sensitive workflows close to the business. Optional hosted models, APIs, email, phone, and business software remain external services when a design uses them.
Human approvals can sit before consequential actions. Execution records can show what ran and what the system observed. Neither feature guarantees correctness, authorization, security, or business success.
07 / PUBLIC ENGINEERING PROOF
Inspect the work
behind the offer.
Selected public systems demonstrate implementation scope across agent memory, orchestration, evidence, compression, and local-first infrastructure.
Mnemes and semantic-memory
A Rust-first memory engine, MCP access layer, agent integration kits, and self-hosted Mnemes server with explicit ownership boundaries.
Inspect semantic-memory ↗Hermes integration path
A RecursiveIntell-enhanced Hermes path integrating Rust services, memory tooling, MCP installation, and multi-agent orchestration.
View the public interaction ↗ClaimLedger and agent-graph
Public Rust components for claim/evidence history, typed graph execution, receipt-bearing tool paths, and bounded orchestration.
Inspect claim-ledger ↗August 5, 2026 · HERMES COMMUNITYTeknium highlighted
Teknium highlighted
Josh’s Hermes work.
Teknium, creator of Hermes Agent, publicly highlighted Josh’s demonstration of a RecursiveIntell-enhanced Hermes setup.
This is a public interaction around Josh’s engineering work, not a customer testimonial, partnership, or product endorsement.
See the public interaction ↗08 / FOUNDER + ENGINEER
One accountable
technical owner.
Josh Stevenson is a systems engineer and founder/operator of RecursiveIntell. He builds local-first AI, agent infrastructure, memory systems, Rust libraries, and operator-grade workflows in public.
RecursiveIntell is a founder-led applied R&D studio and public engineering portfolio. Public source shows breadth and implementation effort; it does not imply a team, customers, funding, compliance, or production suitability.
About Josh + RecursiveIntell →YOUR FIRST MOVE
Tell me what your
business keeps
doing by hand.
A sentence is enough to start. I’ll help determine whether the right next move is a workflow map, a pilot, a knowledge build, or focused technical consulting.