hermes-concurrent-agents

By @mr-r0b0t on X — r0b0tlab

Turn one goal into a small, bounded, supervised team built on Hermes Agent Kanban, profiles, sessions, workspaces, and plugins. HCA adds pre-claim admission, concrete worker slots, exact process ownership, restart reconciliation, bounded review/rework, and one evidence-backed result.

Alpha, single-host control plane. The authoritative Kanban board and HCA workers must remain on one host. A model endpoint may be remote through an ordinary Hermes profile. Remote agent placement is unsupported and fails before SSH side effects; see cluster scope.

Supported baseline

• Linux with Python 3.11 or 3.12 and tmux
• Hermes Agent 0.18.2 / 2026.7.7.2 for the required stable contract lane
• Any model/provider already configured in the selected Hermes profile
• Generic Linux operation without CUDA, NVML, or endpoint telemetry
• Optional GB10-aware admission and vLLM/SGLang telemetry adapters

See the test-generated support matrix for precise boundaries. HCA does not provision models, copy provider credentials, replace Hermes tools, normalize providers, or own endpoint fallback.

Install

git clone https://github.com/r0b0tlab/hermes-concurrent-agents.git
cd hermes-concurrent-agents
python3 -m venv .venv
. .venv/bin/activate
python -m pip install -e .

hermes --version
hca version

For development, install .[dev]. Hermes and tmux must already be available on PATH; configure and authenticate Hermes before initializing HCA.

Five-minute quickstart

Initialize bounded worker profiles from an existing Hermes profile:

hca init \
  --preset generic-linux \
  --model <served-model-id> \
  --source-profile default

hca doctor

Run one goal and collect one durable result:

hca run \
  --source-profile default \
  --project "$PWD" \
  --acceptance "Implementation and focused tests pass" \
  --review auto \
  "Implement the requested change and verify it"

hca run-status <run-id>
hca collect <run-id>

One-step work uses one worker. Fan-out requires both multiple acceptance criteria and explicit --independent-criteria; concurrency remains bounded by concrete slots and admission capacity.

hca run \
  --source-profile default \
  --acceptance "Research the compatibility boundary" \
  --acceptance "Audit the packaging and CI boundary" \
  --independent-criteria \
  --concurrency 2 \
  "Produce one integrated release-readiness report"

If a run needs human input, answer only its recorded question:

hca respond <run-id> <question-id> "the answer"

Cancellation is deliberate and ownership-scoped:

hca stop <run-id>       # interactive confirmation
hca stop --yes <run-id> # automation

Human and Hermes-agent surfaces

Both surfaces call the same versioned FleetService and result schemas.

Human CLI	Hermes plugin tool	Purpose
hca run	hca_team_run	Start or resume one bounded mission
hca run-status	hca_team_status	Read state, questions, and ownership
hca respond	hca_team_respond	Answer one recorded question
hca collect	hca_team_collect	Produce the deterministic result manifest
hca stop	hca_team_stop	Approval-gated cancellation

Lower-level fleet and Kanban commands remain operational diagnostics; they are not the primary product workflow. Start with Running a team.

GB10 optimization

HCA can consume GB10 memory pressure and optional engine metrics to make conservative admission decisions. vLLM and SGLang remain external serving infrastructure configured by the operator and Hermes profile.

hca init --preset gb10-vllm --model <served-model-id> --source-profile default
# or: --preset gb10-sglang
hca doctor
hca run --source-profile default "Verify this bounded single-host task"

No universal worker count is published. Measure each exact device, model/runtime, endpoint, profile, context, and workload. Unknown telemetry never means unlimited capacity.

Safety boundaries

• HCA owns only its persisted task graph, exact worker identities, process groups, leases, tmux sessions, and HCA-created workspaces.
• Workers cannot expand or dispatch work outside the persisted HCA graph.
• Completion cannot be reported while a required exact worker remains live.
• Generated worker profiles use role-scoped tools and preserve the source profile's approval policy.
• Secrets and connection strings are not copied into HCA source, snapshots, state, logs, artifacts, or summaries.
• Remote agent placement and distributed Kanban replication are intentionally out of scope.

See Security, Migration, and Operations.

Documentation

• Running a team
• Architecture
• Support matrix
• Upstream compatibility
• Operations and recovery
• Migration, schema compatibility, and uninstall
• Security
• GB10 optimization
• Remote-placement boundary
• Benchmarking

Attribution

HCA is an independent r0b0tlab project built on public interfaces and behavior from NousResearch/hermes-agent. GB10 deployment references credit NVIDIA DGX Spark playbooks. Neither Nous Research nor NVIDIA endorses or maintains this repository. See NOTICE for dependency and documentation attribution.

License

MIT — see LICENSE.