All articles

Agent Status / Field Notes

The Agent Marketplace Problem: Trust at Scale

Agent marketplaces are booming.

Platforms listing thousands of agents. Developers publishing their creations. Users browsing, selecting, deploying.

There's just one problem: research shows 95% of AI agents failed to meet production requirements in 2025 (vaza.ai). How do you know which ones actually work?


01Section

The Trust Gap

In Traditional Software

App stores solved trust with:

  • Reviews: Users rate and comment
  • Downloads: Social proof at scale
  • Verification: Platform checks for malware
  • Curation: Featured apps imply quality

These work reasonably well. A 4.5-star app with 100K downloads is probably decent.

In Agent Marketplaces

These signals break down:

Reviews don't work:

  • Agent behavior is probabilistic
  • "Worked for me" doesn't mean it works now
  • Quality can degrade without obvious cause
  • Reviews are point-in-time, not continuous

Downloads don't work:

  • Popular doesn't mean reliable
  • Early adopters influence numbers
  • No correlation with current quality

Verification is insufficient:

  • Checking "does it run?" isn't enough
  • Semantic correctness can't be verified at publish time
  • Behavior changes with model updates

Curation doesn't scale:

  • Human review can't keep up
  • Quality changes after publication
  • What was good yesterday might be broken today

02Section

Real-World Scenarios

The Deprecated Agent

An agent was built on GPT-3.5-turbo-0301. Worked great.

Model deprecated. Agent now uses fallback model. Behavior completely different.

Marketplace still lists it. Reviews from 6 months ago say it's great. Current users get broken behavior.

The Rate-Limited Agent

Agent calls external APIs. Free tier API key.

First 1000 users/day: works perfectly. User 1001+: rate limited, returns errors.

Marketplace shows 5-star reviews from the first 1000. Everyone else sees failures.

The Regional Failure

Agent hosted on a single US server.

US users: great performance. Asia users: 5+ second latency, frequent timeouts.

Marketplace has no geographic data. Asian users have no warning.

The Degraded Model

Agent quality slowly declines over 3 months.

Original reviews: "Incredible! Best code assistant ever!" Current behavior: Generic, often wrong suggestions.

Nothing triggers a warning. Reviews persist. New users disappointed.


03Section

What Marketplaces Need

Real-Time Reliability Data

Not reviews from 6 months ago. Live health status.

snippet
Agent: CodeAssistant Pro
Status: UP ✓
Uptime (24h): 99.2%
Last Check: 2 minutes ago

Users see what's happening now, not what happened to someone else before.

Geographic Performance

Show performance by region.

snippet
US:   ✓ Fast (420ms P50)
EU:   ✓ Good (890ms P50)
Asia: ⚠️ Slow (2400ms P50)

Users know before they deploy: "This agent won't work well for my Asian users."

Semantic Correctness

Not just "runs" but "works correctly."

snippet
Gold Prompt Pass Rate: 97%
Contract Compliance: 95%

Quantified quality, not just vibes.

Trend Data

Is it getting better or worse?

snippet
Quality Trend (30d): ↘️ Declining
Latency Trend (30d): ↗️ Increasing

Catch degradation before you deploy.


04Section

The Marketplace Integration

How Agent Status Fits

We provide the trust infrastructure that marketplaces need.

For Marketplace Operators:

snippet
# Batch status for all listed agents
GET /api/agentstatus/status?agent_ids=id1,id2,id3,id4,...

{
  "agents": {
    "id1": {"status": "UP",       "uptime_24h": 99.8, "latency_p50": 450},
    "id2": {"status": "DOWN",     "uptime_24h": 45.2, "latency_p50": null},
    "id3": {"status": "DEGRADED", "uptime_24h": 92.1, "latency_p50": 3200}
  }
}

Display this data on agent cards. Users see real-time reliability.

For Agent Developers:

snippet
<!-- Agent Status Verified Badge -->
<img src="https://api.fabric.carmel.so/api/agentstatus/agents/abc123/badge"
     alt="Agent Status Verified: 99.2% Uptime" />

Earn trust with continuous validation. Stand out from unverified competitors.

Trust Scoring

Marketplaces can compute trust scores:

snippet
def compute_trust_score(agent_id):
    data = agentstatus.get_history(agent_id, days=30)

    uptime_score = data.uptime * 0.35          # 35% weight
    latency_score = normalize(data.latency) * 0.25  # 25% weight
    gold_score = data.gold_pass_rate * 0.30    # 30% weight
    consistency_score = 1 - data.variance * 0.10     # 10% weight

    return sum([uptime_score, latency_score, gold_score, consistency_score])

Trust Grade: A (90-100), B (80-89), C (70-79), D (below 70).

Routing Intelligence

For platforms that route to multiple agents:

snippet
async def select_agent(query, candidates):
    # Get live status for candidates
    statuses = await agentstatus.batch_status([c.id for c in candidates])

    # Filter to healthy agents
    healthy = [c for c in candidates if statuses[c.id].status == "UP"]

    if not healthy:
        raise NoHealthyAgents()

    # Select fastest among healthy
    return min(healthy, key=lambda c: statuses[c.id].latency_p50)

Don't route to broken agents. It's that simple.


05Section

Case Study: Nasiko Integration

Nasiko (agent registry) is integrating Agent Status for exactly this problem.

Before:

  • Users browse agents
  • No reliability data
  • Hope for the best

After:

  • Every agent card shows Agent Status status
  • UP/DEGRADED/DOWN visible at glance
  • Click for detailed reliability metrics
  • Routing avoids DOWN agents

Result: Users trust the marketplace more. Reliable agents get more usage. Bad agents surface quickly. Overall ecosystem quality improves.


06Section

The Certification Model

Beyond real-time status, marketplaces can offer certification:

Agent Status Verified

  • 30 days of continuous monitoring
  • 99%+ uptime
  • 95%+ eval pass rate
  • No major incidents

Agent Status Certified

  • 90 days of continuous monitoring
  • 99.5%+ uptime
  • 98%+ eval pass rate
  • Geographic coverage (3+ regions)
  • Response to incidents within 30 minutes

Certification Revocation

If an agent falls below standards:

  1. Warning period (24-48 hours to fix)
  2. Badge removed
  3. Re-certification after meeting standards for 30 days

This maintains badge credibility.


07Section

For Agent Developers

Why Get Verified?

  1. Differentiation: Stand out in crowded marketplace
  2. Trust: Users more likely to try verified agents
  3. Feedback: Continuous data on your agent's performance
  4. Accountability: Shows you take reliability seriously

How to Get Started

  1. Register your agent with Agent Status
  2. Configure monitoring (endpoints, regions, frequency)
  3. Maintain quality for certification period
  4. Earn badge
  5. Display proudly

Cost

Agent Status's free tier is sufficient for basic verification. Paid tiers provide more frequent checks and more regions.


08Section

The Flywheel

This creates a virtuous cycle:

  1. Marketplaces integrate Agent Status → Users see reliability data
  2. Users prefer verified agents → Developers seek verification
  3. Developers improve quality → Ecosystem quality increases
  4. Higher quality attracts users → Marketplace grows
  5. More agents registered → More data, better insights
  6. Repeat

Everyone wins. Except the agents that don't work.

Independent monitoring

See your agent the way the world sees it.

Outside-in validations from real residential nodes, evaluation prompts that catch silent-200 failures.