Nomos

Multi-agent orchestration for work that must reach a verified end state

Long context agentic coordination | End-state reasoning | Integrated tool calling | Durable self-correction and self-improvement

Nomos turns broad objectives into coordinated, evidence-backed operations. It is built for work that crosses people, agents, tools, repositories, documents, and long-running processes, where a useful system must do more than produce a draft.

Nomos works backward from the required end state. It resolves intent, identifies missing evidence and operational dependencies, coordinates independent work in parallel, reviews actions before execution, learns from observed outcomes, and keeps the objective open until the real result is validated.

From AI assistants to coordinated operations

Most assistants optimize one response. Most automation platforms require a human to define every step in advance. Nomos sits between those approaches: the model owns a dynamic objective graph while the runtime provides durable execution, identity, evidence, correction, and continuation boundaries.

Business need	Nomos capability	Operational result
AI transformation	Converts open-ended requests into model-owned objective graphs, actions, evidence, and verified completion	Move from chat-only pilots to delegated work that can be reviewed, resumed, and measured
Multi-agent operations	Coordinates drafting, research, reproduction, experimentation, triage, patching, validation, and reporting workers	Run independent work concurrently and reconcile it into one accountable outcome
On-demand digital teams	Creates task-specific workers and evidence dependencies from the objective rather than a fixed persona list	Assemble the right working group for each incident, project, investigation, or customer request
Long-running automation	Preserves task graphs, workspace state, tool observations, receipts, and learned corrections	Resume work after process restarts or handoffs without rebuilding context from a summary
Engineering execution	Connects repository inspection, reproduction, repair, tests, regression review, and handoff	Move from issue report to verified change in one continuous evidence chain
Research and analysis	Coordinates source collection, experiments, comparison, challenge, synthesis, and disclosure	Produce decisions that retain their evidence trail instead of collapsing research into an unsupported answer
Operational recovery	Uses current state, runbooks, tools, confidence, and validation edges to adapt when the environment changes	Reduce repeated failure loops and keep recovery tied to observed system state
Agent platform consolidation	Exposes one orchestration runtime through CLI, HTTP, OpenAI-compatible tools, MCP, A2A, and Docker	Give multiple applications and model clients a common long-horizon operating layer
Continuous capability improvement	Correlates verified outcomes with the exact model route, reviewer, workers, and operational state that produced them	Improve future execution without mixing users, projects, or unrelated task state
Governed autonomy	Applies schemas, authentication, workspace containment, signed observations, approvals, and evidence-bound completion	Increase autonomy while retaining operational control and auditability

What makes Nomos different

Graph end-state reasoning

Nomos begins with the state that must be true when the work is finished. It then works backward through intent, confidence, operations, evidence, dependencies, and validation. The graph can expand when execution discovers a new requirement; it is not constrained to a static checklist or a fixed number of reasoning steps.

Native multi-agent coordination

Worker roles are capabilities in the active graph, not host-authored role-play. Nomos can coordinate independent work concurrently and pass actual evidence identities to dependent workers.

Common work products include:

• Drafting for candidate actions, plans, responses, and disclosures.
• Research for source discovery, comparison, and synthesis.
• Reproduction for isolating failures across environments and variants.
• Experimentation for resolving uncertain causal edges.
• Triage for impact, priority, ownership, and competing hypotheses.
• Patching for minimal evidence-grounded repairs.
• Validation for regression checks, edge cases, and end-state proof.
• Disclosure for clear user, customer, operational, or executive reporting.
• Coordination for worker identity, dependencies, knowledge exchange, and reconciliation.

Nomos records both observed exchanges and learned knowledge transfer. This allows one worker's verified finding to alter later drafting, action selection, and validation without reducing the transfer to an unverified summary.

Integrated tool calling

Tool calling ships inside every complete Nomos profile. The action component is integrated under the profile's model authority and is not a separate product, service, or download.

The integrated action capability proposes semantically relevant function calls and structured arguments. Nomos remains responsible for objective intent, worker coordination, graph fit, confidence, action review, execution authority, observation handling, correction, and final completion.

objective and required end state
              |
      intent + task graph
              |
  parallel workers and evidence
              |
 integrated candidate tool calling
              |
 Nomos review -> execution -> observation
              |
 correction -> validation -> completed outcome

Self-correction during the current task

A failed action is not merely stored for a future session. Nomos binds the signed observation to the exact call, draft, worker, graph edge, and producing route. The active task can then revise arguments, choose another capability, acquire missing evidence, reopen an uncertain decision, or produce a new reviewed draft.

Correction is candidate-specific. A denied action in a parallel batch does not indiscriminately degrade approved sibling actions, and an outcome cannot redirect credit to an unrelated draft.

Verified self-improvement

Nomos improves from real outcomes rather than from ungrounded critique. Verified success or failure can update the exact surfaces that produced the behavior:

• action and response reviewers;
• recurrent confidence and correction routes;
• candidate and operational representations;
• NoNE expert participation and cross-capability transfer;
• model-owned task, traversal, and coordination state;
• worker dependencies and knowledge exchange;
• native proposal fidelity;
• learned reliability for operational capabilities.

If an initial attempt fails, Nomos can correct inside the current task. If a corrected attempt also fails, it can persist improvement before trying again. Learning is session-owned, cold-loadable, and correlated with the producing state, so delayed feedback cannot borrow an unrelated live task.

Long-context, long-horizon continuity

Nomos is configured for a native context window of 4,194,304 tokens. Long context supports large repositories, document collections, logs, histories, and ongoing projects, while durable task state supports work that outlives one context or one server process.

Context processing remains connected to attention, KV state, objective state, intent, action state, NoNE participation, model-owned coordination, and recurrent confidence. Native live analysis can expose token progress and model-owned completion state without replacing the answer with a host-selected shortcut.

Industries and operating teams

Software and platform engineering

Use Nomos to investigate incidents, inspect repositories, reproduce defects, compare repair strategies, implement changes, execute tests, review regressions, and create evidence-linked handoffs. Multiple workers can investigate in parallel while dependent patching and validation wait for grounded findings.

IT, security, and operations

Coordinate state inspection, runbook execution, incident triage, remediation, validation, and status reporting. Nomos can adapt when live state differs from a stale plan and preserve who requested an action, what executed, and which observation changed the response.

Finance and regulated operations

Coordinate document review, policy checks, reconciliations, exception handling, and evidence packages across multiple systems. Session isolation and signed observations make it easier to retain the operational chain behind a decision.

Healthcare administration and life-science operations

Support high-context administrative and research workflows that combine records, procedures, evidence collection, review, and follow-up. Nomos is suited to coordinating operational work and human review; deployments remain responsible for domain-specific authorization and safety controls.

Manufacturing and field operations

Connect telemetry checks, maintenance evidence, incident reproduction, repair planning, validation, and status consolidation. Failed operations reduce learned capability confidence instead of being misreported as successful merely because they ran.

Research, legal, and professional services

Run parallel source collection, fact comparison, challenge, drafting, review, and disclosure. The objective graph preserves unresolved questions and evidence dependencies until the final work product is ready for its intended audience.

Capabilities by use case

Use case	What Nomos provides
Complex project agent	Dynamic objectives, parallel work, dependencies, durable state, correction, and end-state validation
Software-delivery agent team	Reproduction, experiment, triage, patch, test, review, disclosure, and exact evidence transfer
Research organization	Long-context evidence gathering, source comparison, challenge workers, synthesis, and grounded reporting
Incident command agent	Live-state inspection, concurrent investigation, remediation planning, execution evidence, and recovery validation
Agentic customer operations	Case state, tool-backed actions, exception correction, escalation evidence, and consistent final communication
Compliance workflow	Structured operations, signed observations, review steps, session isolation, and durable receipts
Model-fleet orchestrator	OpenAI-compatible, MCP, A2A, HTTP, and CLI boundaries around one persistent objective and worker graph
Persistent workspace agent	Repository and file operations, resumable tasks, artifacts, corrections, learned outcomes, and cold continuation
Tool-rich application backend	Integrated function-call candidates plus Nomos intent, confidence, review, execution, and follow-up orchestration
Continuous improvement system	Outcome-correlated learning for routes, reviewers, workers, action candidates, and capability reliability

What you can build

• Autonomous engineering teams that move from report to verified patch.
• Research agents that split an investigation, reconcile evidence, and retain provenance.
• Incident-response coordinators that adapt to changing infrastructure state.
• Persistent project agents that resume after interruption with the same objective graph.
• Internal operations agents that coordinate tools, approvals, artifacts, and human input.
• Agent platforms that expose one durable orchestration layer to multiple clients and models.
• Review and disclosure workflows that connect the final document to the work that produced it.
• Self-correcting tool workflows that learn from authenticated execution outcomes.

Choose your Nomos

Nomos uses the same complete-profile family structure as Nexum. A profile is a standalone model, runtime, server, CLI, Docker definition, and operating guide. Profiles are never hand-selected tensor subsets.

Profile	Availability	Operating emphasis
Nomos-Universal/	Available	Balanced multi-agent orchestration, long-context work, integrated actions, durable correction, and continual improvement
Nomos-Lite/	Family slot	Responsive orchestration for narrower local and operational deployments; published only when a complete validated profile exists
Nomos-Expanded/	Family slot	Broader concurrent work, deeper long-running objectives, and expanded capability state; published only when complete

The current release is Nomos-Universal. Future profiles can be added as complete sibling folders without changing the root integration contract.

Repository layout

Nomos/
├── README.md
├── config.json
├── params.json
├── meta.yaml
├── release_family.json
├── assets/
├── containers/
└── Nomos-Universal/
    ├── README.md
    ├── config.json
    ├── tokenizer.json
    ├── chat_template.jinja
    ├── model/
    │   ├── safetensors/
    │   ├── tensor_map.json
    │   └── components/tool_calling/
    ├── runtime/
    ├── containers/
    └── release.json

model/components/tool_calling/ is an internal part of the Nomos model profile. It is loaded, reviewed, and governed by Nomos and is not an independent release.

Download and validate

hf download namenotfoundai/Nomos \
  --include "Nomos-Universal/**" \
  --local-dir ./Nomos

python ./Nomos/Nomos-Universal/runtime/verify_release.py

CLI

Install the compiled runtime shipped with the profile:

python -m pip install \
  ./Nomos/Nomos-Universal/runtime/dist/nomos_runtime-0.3.0-cp312-cp312-linux_x86_64.whl

nomos --help
nomos doctor
nomos bundle validate --model ./Nomos/Nomos-Universal --deep
nomos prompt --model ./Nomos/Nomos-Universal \
  --prompt "Investigate the objective, coordinate the work, execute the required actions, and verify completion."
nomos chat --model ./Nomos/Nomos-Universal

The same entrypoint exposes tools, operations, durable harness sessions, correction, improvement, outcomes, receipts, and the HTTP server.

Server

export NOMOS_API_KEY=replace-me
nomos serve \
  --host 0.0.0.0 \
  --port 8000 \
  --model ./Nomos/Nomos-Universal

Nomos provides an authenticated OpenAI-compatible chat and tool boundary while preserving its own objective, worker, action, correction, and learning state.

Containers

Build the current complete profile:

docker build \
  -f Nomos-Universal/containers/Dockerfile \
  -t namenotfoundai/nomos:universal \
  Nomos-Universal

With no command override, the image runs the server. Supplying arguments runs the CLI:

docker run --rm --gpus all \
  -e NOMOS_API_KEY=replace-me \
  -p 8000:8000 \
  namenotfoundai/nomos:universal

docker run --rm --gpus all \
  namenotfoundai/nomos:universal doctor

Security and operational control

Nomos includes authenticated service boundaries, session namespaces, strict tool schemas, workspace containment, signed observations, call correlation, secret redaction, durable receipts, and exact outcome association. Organizations remain responsible for network policy, identity integration, tool permissions, human approval rules, and domain-specific safeguards around deployed actions.

Citation

@software{nomos2026,
  title = {Nomos: Multi-Agent Graph End-State Orchestration},
  author = {Name Not Found AI},
  year = {2026},
  url = {https://huggingface.co/namenotfoundai/Nomos}
}