In this publication
Key takeaway
Route each mission to the smallest useful set of specialists. Keep permission enforcement, evaluation, and recovery explicit at every handoff.
The design problem
Adding skills and agents creates more possible handoffs, competing instructions, and operational dependencies. The GovDOSS Adaptive Expert Mesh proposes a way to organize that work around a defined mission, reusable expertise, and bounded execution.
The July 2026 design separates a skill—a reusable method or set of instructions—from an agent, which acts with a specific objective, tools, state, and limits. This makes skill reuse possible without assigning every skill its own autonomous actor.
Skills and agents play different roles
Reusable skills can serve more than one bounded agent. The example cells show a relationship, not deployed instances.
Source: GovDOSS Adaptive Expert Mesh · proposed architecture v0.1 · July 23, 2026.
Read figure description
Reusable methods and instructions are bundled into or shared by expert cells. Each cell is a bounded agent with an objective, tools, state and limits. Expert cells return their results to the mission coordinator. A skill does not require its own autonomous agent.
How the proposed architecture works
A mission interface captures the requested outcome and constraints. A code-controlled router selects relevant expert groups. Specialists return structured results to a central coordinator, while separate evaluation checks the evidence and action boundaries.
- Mission context: the objective, sources, assumptions, permitted data and tools, budgets, and stop conditions.
- Specialist work: focused groups for decisions, assurance, security and AI, commercialization, operations and knowledge, and R&D.
- Evaluation: checks for evidence quality, contradictory findings, security boundaries, and task-specific acceptance criteria.
- Execution and recovery: authorized operations with observable results, exception handling, and a defined rollback path.
The proposed mission architecture
Logical responsibilities and handoffs. Assurance findings do not grant execution permission; authorized actions still require applicable approval.
Source: GovDOSS Adaptive Expert Mesh · proposed architecture v0.1 · July 23, 2026.
Read figure description
The mission interface captures objectives and constraints and owns final synthesis. A deterministic router builds the mission envelope and selects relevant cells from Governance and Decision, Assurance and Compliance, Security and Regulated AI, Growth and Commercialization, Operations and Knowledge, and Evolution and R&D. Structured specialist results return to the coordinator. Independent assurance checks evidence, contradictions and action boundaries. Sandboxed workers act only within authority and applicable approval. Memory, observability and evolution capture traces and outcomes to inform future proposals. This is a logical view, not a deployed network topology.
Authority travels with the work
A handoff should preserve who is acting, what resource is in scope, how identity is established, which permissions and approvals apply, and what action is allowed. The design applies the GovDOSS SOA⁴™ control chain to those questions.
A specialist may recommend a change without having permission to execute it. Changes to permissions, production behavior, and the system’s own operating methods need a distinct release decision. The proposed architecture does not allow agents to expand their own authority.
The SOA⁴ control chain at a handoff
Read the chain in order. An unresolved identity, permission or approval requirement leads to stopping or clarification before action.
Source: GovDOSS Adaptive Expert Mesh · proposed architecture v0.1 · July 23, 2026.
Read figure description
Subject: who is acting? Object: what is in scope? Authentication: is identity established? Authorization: is the action permitted? Approval: is required approval present? Action: execute only within bounds. At any unresolved requirement, stop, clarify or seek applicable approval. Recommendations do not expand the agent’s authority.
Learn through bounded experiments
The design uses an OODA cycle to connect observation, mission framing, selection, action, and feedback. Candidate improvements remain proposals until their evidence supports promotion.
The proposed pilot measures routing quality, evidence-supported claims, human corrections, cost and latency per mission, failure recovery, and rollback success. These are planned evaluation measures; this brief reports no observed improvements against them.
From observation to a proposed improvement
The source extends OODA with verification and learning. An improvement remains a proposal until a separate promotion decision.
Source: GovDOSS Adaptive Expert Mesh · proposed architecture v0.1 · July 23, 2026.
Read figure description
Observe evidence, orient to the mission, decide on bounded work, act within authority, verify actual outcomes, learn from the result, evolve by proposing an improvement, and re-observe new evidence. Repeat the cycle. Promotion is a separate governed decision. The brief reports no measured pilot outcomes.
Limitations & next step
This brief summarizes a proposed architecture, version 0.1. It is not evidence of a deployed platform, completed pilot, independently validated performance, or a compliance determination.
Next research step
Define the handoff schemas and acceptance thresholds, then compare a bounded pilot against the existing skill-based workflow before any production promotion.
Evidence needed before promotion
Only the architecture proposal is established by this brief. Dashed stages summarize proposed work; they do not indicate completion or a measured maturity level.
Source: GovDOSS Adaptive Expert Mesh · proposed architecture v0.1 · July 23, 2026.
Read figure description
Current evidence is architecture proposal version 0.1. Next define pilot handoff schemas and acceptance criteria. Then compare a bounded pilot against the existing workflow. Independent validation and release authority are required before promotion. No pilot results or acceptance thresholds are supplied here.
Questions about this research? Contact GovDOSS.