Run Graph
Watch every step, branch, and retry as it happens. Click any node to see exactly what went in and what came out.
step_04 · kb.search
- tool
- kb.search
- attempt
- 2 of 3
- latency
- 842ms
- tokens
- 1,204
- status
- resolved
Agents that finish what they start.
Apkar gives every run a plan, a retry policy, and a full audit trail — so your automation survives contact with production.
- p50 overhead
- 12ms
- retention
- 90 days
- regions
- 14
run_8f2a1cCompleted
Running in production at
- NORTHWIND
- Kestrel
- LUMEN BAY
- Foundry
- HALDEN
- Arcus
- VOSS
- Meridian
4.8 / 5214 reviews
“We deleted 4,000 lines of retry logic and replaced it with one policy file.”
The real failure mode
Most agents don't fail because the model was wrong. They fail because nothing was watching.
A run stalls on a rate-limited tool call at step nine. Nobody notices for two hours. There is no record of what the agent had already done, so the only safe move is to start it over from scratch.
The model was never the problem. The missing pieces were a checkpoint, a retry rule, and somewhere to look afterwards.
- Completed
- Stalled at step 9
- Recovered on retry
Built for production
Everything a run needs to be trusted.
Retry Policies
Per-tool exponential backoff, declared in one file instead of scattered through your handlers.
Audit Trail
Every prompt, tool call, and token — retained 90 days and replayable.
Human in the Loop
Pause a run, review the state, approve, and resume from the same step.
Cost Guards
A hard ceiling per run. The agent stops. It does not quietly overspend.
How it works
Three steps, no glue code.
01
Define
Describe the agent in TypeScript or the visual builder. Both compile to the same graph.
agent("triage")
.step("kb.search")
.retry({ max: 3 })02
Run
Parallel execution with a checkpoint written after every tool call.
01
02
03
04
03
Replay
Rerun any past run step by step, with the exact inputs it saw the first time.
step 06 / 092.8s
Build vs buy
What you skip by not building it yourself.
| Rolling your own | With Apkar | |
|---|---|---|
| First production agent | ~6 weeks | 2 days |
| Retry & backoff logic | ~4,000 LOC to maintain | One policy file |
| Step-level replay | Build it yourself | Built in |
| Per-run cost ceiling | Manual accounting | Hard limit |
| SOC 2 evidence | Your problem | Exportable |
Integrations
Your stack, already wired.
Tools your agents can call on day one, with auth and rate limits handled.
and 64 more
34%
faster support resolution in one quarter
Northwind moved 11 support workflows to Apkar and cut median handle time from 6m 40s to 4m 24s.
Read the case studyYour first agent runs in under ten minutes.
Start building — free
No credit card · 10,000 free steps