Named-user seats
Measure utilization and identify stranded spend. Keep the entitlement private to its assigned user. Reassign or change seat tiers through provider-supported admin controls instead of sharing credentials.
A2Agent gives companies one control plane for named-user AI seats, organization-level credits, enterprise usage, APIs, and local compute. It does not pool employee passwords or pretend a per-user allowance is transferable.
Credential-local · entitlement-aware · budget/model policy · worker trust · observe-first execution evidence
Some employees barely touch coding tools. Other teams hit limits constantly. The wrong solution is sharing logins. The right solution is separating seat management from provider-approved shared capacity.
Measure utilization and identify stranded spend. Keep the entitlement private to its assigned user. Reassign or change seat tiers through provider-supported admin controls instead of sharing credentials.
Route eligible workloads across organization-owned OpenAI/Anthropic API workers when the company has the right execution contract. Host managers keep credentials local with Windows DPAPI or macOS Keychain storage.
Use company GPUs and local models through the fixed-loopback Ollama worker. A Capacity Grant keeps internal sharing rights explicit and revocable.
A2Agent treats routing rights as data, not a marketing assumption. Every organization job must pass entitlement, workspace and execution policy before it can enter the capacity queue.
The router already enforces adapter, provider, model, scope, budget, entitlement, availability and concurrency boundaries. Worker v5 adds capability evidence, and capacity_execution_v1 now records how the worker that actually leased the job looked at that moment — without changing selection yet.
worker_protocol = 5
trust_enforcement = observe
evidence_score = capacity_execution_v1
affects_selection = false
worker_pull_leasing = true
public_pool = false
billing_enforced = falseThe downloadable worker reports a bounded capability envelope so A2Agent can show whether the local runtime passed a recent adapter-specific probe — without collecting provider credentials.
Worker/protocol version, OS and architecture, runtime version, model pinning, fixed-endpoint posture and adapter-specific readiness checks. Secret values and prompt/result content stay out of the trust envelope.
The Console distinguishes legacy unknown, self_reported, and freshly probed workers. verified is reserved for future stronger platform/provider certification and cannot be self-assigned.
Successful attempts, failed attempts, retryable failures, runtime EWMA, capacity headroom, shadow-price fit and trust health now produce a bounded operational score snapshot.
New workers do not start at 100% reliability, missing runtime telemetry is not treated as perfect, and every component has a visible versioned definition.
Confidence-smoothed success vs failed execution attempts.
EWMA evaluated against quick / standard / heavy workload baselines.
Worker minimum shadow value versus current job shadow value.
Available concurrency before the current claim.
Fresh probe state without overriding rights or company policy.
Provider OAuth/session credentials stay on the machine and remain tied to the named user or approved organization service identity.
A2Agent does not rotate named-user seats to circumvent per-user limits. If a provider says an allowance is per member, A2Agent treats it as per member.
The score is observe-only until real telemetry is large and stable enough to justify explicit eligibility thresholds. Global central dispatch is not implemented.