Hermes Agent
The self-evolving digital employee that learns once and executes forever
A persistent autonomous agent built on the Hermes framework. Operates across Slack, Telegram, WhatsApp, Discord, and CLI with a built-in learning loop that writes its own reusable Python skills. Every problem it solves becomes a permanent capability — zero-maintenance scaling without ongoing developer babysitting.
Overview
Hermes is the always-on operational layer of the Hannah Imagineer ecosystem. Hand it a recurring task — reporting, research, triage, scheduled automations — and it builds the skill on first run, persists it, and re-uses it forever after. Skills compose: every new RPC becomes a primitive the orchestrator (and other Hermes instances) can call. The deployment footprint scales from a $5 VPS to a multi-GPU cluster without code changes.
Primitives
Persistent memory across sessions and platforms
Automated skill creation — agent writes its own reusable Python RPCs
Containerized parallel sub-agents for delegated workstreams
Native multi-platform gateway (Slack, Telegram, WhatsApp, Discord, CLI)
Natural language cron scheduling for unattended automation
Full web browser control and autonomous research
Deploys to local server, $5 VPS, GPU cluster, or serverless
Outcomes
- 01Replace a junior ops headcount with an always-on operator
- 02Compound capability — every solved task becomes a permanent skill
- 03Zero-maintenance scaling: no developer babysitting after deploy
- 04Unified surface: one operator across every chat tool the team uses
Integrations
Autonomy & guardrails
- Payment or fund movement over configured threshold
- Production deploys and infra mutations
- Outbound communications to external customers
- Deletion of user data or repository history
Guardrails & requirements
Guardrails
- Skill execution sandboxed per container
- Human approval required for destructive shell or financial actions
- Egress-policy enforcement on all autonomous web sessions
- Audit log of every skill invocation with full reasoning trace
Requirements
- Docker or container runtime on host
- One LLM API key (Anthropic, OpenAI, or compatible)
- Persistent volume for memory store (~5GB recommended)
Technical specifications
Runtime
- Harness
- Hermes Framework
- Deployment
- Self-hosted (Docker) · Managed cloud · Hybrid
- Data residency
- Customer-controlled (self-host) or US/EU regions (managed)
- License
- Commercial · Apache 2.0 core
- Version
- v1.2
Models & tooling
- Models
- claude-opus-4-7claude-sonnet-4-6gpt-4olocal: llama-3.1-70b
- Tooling
- python-rpc-skillsbrowser-controlcron-schedulercontainer-spawnvector-memory
Reliability & limits
- Latency SLA
- ~1.2s p50 / ~4.5s p95 per turn
- Rate limits
- 60 turns/min per instance · 10k tasks/day soft cap
Security & compliance
- Auth model
- OAuth2 / API key / per-platform bot tokens
- Compliance
- SOC 2 Type II (managed tier)GDPREU AI Act compliantCustomer-managed encryption keys
Architecture notes
Memory architecture
Three-tier memory store:
- Hot — in-context working set (last N turns + active skill state)
- Warm — per-platform conversation history in Postgres
- Cold — vector store (1536-dim) of every solved task, indexed by problem signature for skill re-use
Context strategy
Hermes assembles context per call: persona + active platform adapter + retrieved skills matching the current intent + last N turns. Skill descriptions are embedded; retrieval is signature-similarity over the problem statement. Context budget is enforced per-model.
Evaluation
Each skill ships with a generated regression case captured from the first successful run. A nightly eval suite replays every skill in a sandboxed container and flags drift. Failure rate, mean tokens-per-task, and skill-reuse-ratio are emitted to the operator dashboard.