Which artificial capability deserves the work?
Compare generations in context, validate performance differences and produce evidence-backed allocation decisions.
Agent management solves the wrong question.
Registries answer how many agents exist. They do not answer which one deserves the work — or how reliably you can prove it.
P1
Agent sprawl
A registry can tell you how many agents exist. It cannot tell you which one deserves the work — that is the wedge.
P2
The new-model paradox
A benchmark-strong new model can still be less reliable than a proven production agent in one specific domain.
P3
Evaluation fragmentation
Eval results, traces, human review notes, business outcomes and cost data live in separate systems that never get reconciled.
P4
The causal attribution gap
A failure can originate in the model, the genome, memory, tools, data, authority, task difficulty, orchestration, runtime or the environment — rarely just “the agent.”
P5
Silent model drift
Provider behavior changes or alias-based endpoints can quietly blur an agent's cognitive origin without anyone noticing.
P6
The ROI-metric gap
Token cost or task-completion rate alone never shows the actual business value produced.
Bring your agents. Keep your stack.
Agents built on different models, runtimes and enterprise platforms are admitted without rebuilding them. XCNTRY normalizes what enters — not how it was built.
At least two different source or runtime profiles must be admitted into the same challenge cohort — a single-vendor demo proves nothing about the category claim.
Genome Passport
Every admitted agent receives a passport that records its model identity, instructions, memory architecture, tool contracts and capability claims — the minimum an evolving population needs to compare like with like.
A model upgrade is not an agent update. A new model requires a new genome and a new birth — an existing agent is never silently swapped onto a new model version.
Fixed at birth (genome)
- Base model & exact version
- Tokenizer version
- Core system-instruction identity
- Base memory-architecture version
- Genome ID & lineage
- Birth certificate
Can evolve (phenotype)
- Episodic memory & task history
- Reputation
- Certified skills
- Company membership
- City relationships
- Budget limit & tool access authority
These are phenotype, not genome — they can evolve without ending an agent's identity.
Context Cell
Champion and challenger generations are compared inside a shared Context Cell: the same purpose, capability, situation and authority profile. Fitness only means something relative to a fixed context.
The unglamorous object that makes the whole category claim true: too broad a context masks real differences, too narrow fragments the sample until no comparison is valid.
Trust controls eligibility. Evidence controls decision quality.
Agent trust ladder
XCNTRY-controlled genome, test chain, birth manifest and attestation.
Strong evidence for model/runtime identity, publisher provenance, tool boundary and configuration.
Some genome components are verifiable; some fields rely on vendor declaration.
Behavior is observable; cognitive/configuration identity is only partially verifiable.
Insufficient evidence for source, identity, or execution boundary; production authority is closed.
Trust is not a performance score. High trust does not excuse a weak outcome; low trust adds constraints on authority ceiling and evidence requirements.
Evidence trust ladder
Agent or vendor self-declaration.
An event observed in the gateway or customer environment.
An event correlated with agent, task, purpose and Context Cell.
Evidence carrying independent verification, signature or trusted provenance.
Evidence closed out with a real acceptance, outcome or business result.
Evidence that has passed difficulty, confounder, context and attribution controls.
A large volume of telemetry does not by itself produce Selection-grade evidence — E5 is a methodological acceptance state, not a storage label.
New models create challengers. Evidence creates champions.
Capability: Supplier Risk Intelligence
G7
Champion
- Mature phenotype
- High reliability
- Higher cost
Current share
Recommended share
+ strategic niche
G12
Challenger
- New frontier model
- Lower human attention
- Higher average quality
Current share
Recommended share
G9
Diversity Challenger
- Different provider/runtime
- Strong under high uncertainty
Current share
Recommended share
uncertainty niche
The newest generation can win. The older generation can remain the best specialist. The city preserves evidence, not fashion.
v0.2 recommends allocation and lifecycle actions under human governance. Production enforcement remains gated.
There is no universal best agent.
Fitness belongs to a capability, a situation and an operating context — never to the agent alone.
Agent Trust
Identity, source and execution-configuration assurance.
Evidence Trust
The decision-usability quality of an evidence package.
Reputation Profile
Multi-dimensional operating history over time.
Contextual Fitness
Risk- and cost-adjusted performance inside one specific Context Cell.
The city never collapses these into one universal agent score. It keeps four axes separate on purpose.
Before replacing an agent, understand why it failed.
A weak outcome may come from the model, genome, data, tools, access, authority, task assignment, orchestration or environment. XCNTRY separates performance from attribution before lifecycle decisions are made.
No punishment without attribution.
Ten competing hypotheses, checked before any blame lands on the agent
“The challenger performed better in this comparison” is a valid claim. It alone does not prove the model caused the improvement, or that the result transfers to every context.
A new challenger never starts high.
A new generation earns task share step by step, each step gated by fresh evidence — not by model novelty or a single strong benchmark.
Each step requires
Live Selection Event — Supplier Risk Intelligence
Before: Champion 80% / Challenger 10% / Diversity Challenger 10%
Recommended: Champion 60% / Challenger 25% / Diversity Challenger 15%
Challenger recommendation: ACCEPT. Diversity Challenger recommendation: PARTIAL — reason: external trust level limits high-risk supplier class.
Adapt, specialize, recover, or leave the active population.
Lifecycle is governed by evidence, attribution and policy. Failure can originate from the agent, model, data, tools, authority, orchestration or environment. Selection follows diagnosis.
Human Gate
Recommendations do not execute themselves. Task-share changes, lifecycle transitions and anything irreversible pass through a human gate — v0.2 recommends under human governance; production enforcement remains gated.
Human review ladder
Routine data analysis.
Low-risk content generation.
Expert context evaluation.
Contract, payment, hiring.
Rights, law, strategy.
Security breach, loss of control.
Low-impact adjustments can run automatically within policy bounds. Anything above it stops for a human.
A rejection does not create a failure label. The reason — missing context, legal constraint, hidden dependency, trust gap, expert knowledge — is recorded and becomes a calibration input.
The long-term asset is the history of what works under real organizational complexity.
Every city becomes more informed about which artificial capability configurations create value in its own environment.
Related XCNTRY topics
Follow the entity graph.
Selection Evidence
Evidence that has passed difficulty, attribution, context and confounder controls well enough to support allocation or lifecycle recommendations.
Open topic
Human Constitution
The set of human-reserved domains, legitimacy rules and final accountability boundaries that constrain machine-scale execution.
Open topic
Purpose Realization Kernel
The non-delegable strategic core that turns human purpose into capability requirements, bearer qualification, authority boundaries and evidence conditions.
Open topic