The observability layer for agents in prod.

Know when your agents are broken before your customers do. Real-time visibility while your agents are running. Not after they crash.

Talk to the founders
For long running agents - Unlimited retention - Query via MCP - White labeling
glassflow.dev
Live1d7d30dCustom
Traces 5Spans 11
AllCompletedFailedFrozen
Trace volumeOK · Error
Search traces…
ClaudeCodexLangChain
NameTrace IDErrorsSpansTokensCostLatency
document-processor2h agoa1b2c3d4e5—13142,800$0.843h 14m
sales-outreach-agent4h agob2c3d4e5f6rate-limit831,400$0.1847m
infra-monitor6h agoc3d4e5f6a7timeout31218,600$1.206h 04m
research-agent1d agod4e5f6a7b8—1589,200$0.622h 31m
support-classifier1d agoe5f6a7b8c9—3420$0.0024s
Total traces
39
Avg duration
1h 54m
Token cost
$18.16
Total tokens
3.1M
Top agents
document-processor
14
infra-monitor
9
research-agent
6
support-classifier
4
Cost per model$18.16
claude-sonnet
$8.42
claude-opus
$6.12
codex
$2.88
haiku
$0.74
Error rate11%
Agent workflow map
fetch-config24m
load-context41m
enrich28m
plan38m
codex8m
write6m
ActiveExecutedFailedRecovered
Span plan-execModel ClaudeLatency 38mCost $0.00

Storing in a DB and querying traces is not for prod. Other tools are built for chats.

Logging traces to a database and querying them yourself works only for testing environments and max your first customer but not when you start to scale. An incident goes live, and the run is too long or too broken to reconstruct with SQL while a customer waits. Learn why querying a DB will make you lose customers.

The chat-era tools don't help. They were built for short sessions that finish in seconds, not agents that run for hours, chain dozens of steps, stay always-on, or span multiple frameworks. When you're holding an SLA, “let me query the database and get back to you” is not a plan.

GlassFlow is built for that. Every step is persisted as it happens, and full sessions stay queryable in milliseconds even after days of runtime. It's the first observability layer that fits agents in production, so you catch the incident before your customer does.

What you get

Alongside the standard features, such as a dashboard, alerts, traces and span view, you will get unique functionalities.

Long-running agents

One session across restarts and sub-agents, not thousands of disconnected requests.

→
Agent workflow maplength 38m
24mfetch-config
↘
41mload-context
↗
38mplan-exec
↘
6mwrite-feature
Span plan-execModel ClaudeLatency 38mCost $0.00

Query via MCP

Your coding agent queries traces over MCP and ships the fix. No dashboard.

→
GlassFlow · mcp · claude-codeLIVE
Claude Codeglassflow_query("show me BillingAgent failures from last week")
GlassFlowFound 3 runs with errors. Retrieving span detail…
GlassFlowRoot cause: STRIPE_API_KEY rotated at 09:12, agent started at 09:14. Key not propagated.

White labeling

Embed observability in your own product, under your own domain.

→
Your company
149agent runs
$1.53model cost
Analyst
Ops

Unlimited trace retention

Recent traces in milliseconds; older ones auto-tier to low-cost S3. Nothing deleted.

→
Storage
Events this month 25%
12,480 / 50,000
Storage used
2.4 GB / 10
Retention
7d14d30d
Storage by agent
AgentTracesSizeOldest
document-processor1420.82 GB7 days ago
infra-monitor890.94 GB7 days ago

Thinking about the agent infrastructure gap

Writing about long-running agents, the stack that does not exist yet, and what needs to be built.

Time horizon of software tasks different LLMs can complete 80% of the time

Why we do not trust autonomous agents yet (and it has nothing to do with trust)

It's not that we're unwilling to trust agents. It's that nobody's built the infrastructure that would make it rational.
→
Context rot versus context window chart

Your agent's context window isn't full. It's just full of the wrong things.

Context rot is measurable, and it starts long before you hit any token limit.
→
Email address copied