Loops
The same work runs again
Repeated requests inside one session, usually because the agent re-reads its own output with no completion limit.
Production agent diagnostics
Know which agents are safe to scale.
Give every agent a track record, then scale the ones that earned it.
Read-only, and real. No signup, no setup.
Agent Behavior Investigation
The agent looped. Costile showed why.
The same agent run, in full: root cause, evidence, and the verified improvement after the fix, the diagnostic report your team would see.
High Impact · Resolved with proof
Support-bot context loop incident
Repeated calls caused by context accumulation. The full conversation
history is resent every request without context limits.
Looping Calls
9
Verified Recovery
93.4%
Incident Window
02:47 → 02:52
Max Token Hits
3 / 9
Session Loop
9 requests in the same session, with tokens escalating from 920 to
1,480. Max token hits occurred on 3 of 9 calls.
Repeated calls
9
Retry window
18 min
Max tokens hits
3 / 9
→ Agent re-reads its own outputs without a completion limit.
Issue
- Repeated calls caused by a context accumulation loop.
- Full conversation history resent without context limits.
Root Cause
- Agent re-reads its own outputs.
- No completion limit, so context keeps growing.
Recommended Fix
- Add a
max_turnsguard. - Set
max_turns: 6. - Reset context after tool completion.
Expected outcome
Stops loop
Repeat calls
~85% lower
Recovery state
Verified
Verification
- Marked monitoring after the guard was deployed.
- Observed 4 follow-up requests in the recovery window.
- Follow-up runs stayed below the guardrail threshold.
Workflow state
Resolved
Recovery proof
Verified
Workflow
Support
Workflow
Support automation · Northstar production AI workflows
What Costile looks for
Your agent can work and still waste most of its calls.
A production agent is not a single model call. It is a chain of models, tools, retries and accumulated context, and one wasteful step pulls the steps after it along with it.
The agent still returns the right answer, so nothing in the logs looks broken, and the waste shows up only as a workflow that gets slower and heavier every week. Costile watches live traffic for these behaviours, and for short requests routed to an expensive model when a cheaper one would have answered identically.
Retry cascades
One failure pulls the rest along
A step retries against the same failing condition, and every attempt drags the calls that follow it into the same work.
Error storms
The session runs on failures
Most of a session's calls come back as provider errors, so the agent keeps building on results that never actually arrived.
Max-token cutoffs
Work is truncated, then redone
Calls stop at the token ceiling, which usually means output is being cut short and attempted again from the top.
Detecting the problem is the easy half.
Helicone and LangSmith record what your agent did, and stop there. That leaves your team to work out the cause, decide on a change, and guess whether it helped. Costile carries the same incident through all five steps.
Detect
Costile reads live traffic for the behaviours above, against rules with a named threshold you can see.
Diagnose
Every incident arrives with the cause named and the requests that prove it, so the conclusion can be checked rather than believed.
Recommend
Costile names the specific change to make, such as a
max_turns guard on an agent re-reading its own
context.
Verify
Costile reads the runs that followed the fix and reports whether it held, or whether the agent still needs attention.
Scale
Each cycle adds to the agent's record, which is what tells you the ones that have earned more traffic.
Agent track record
Every agent builds a record. None of them gets a score.
We do not print a number out of a hundred. A score like that only means something against benchmark data across many agents, and nobody has that yet.
Costile reads the history as a direction instead. An agent is safe to scale only when it is healthy now and not degrading, because a healthy agent heading the wrong way is exactly the one you should not give more traffic to. Where a version has too little traffic to support a trend, Costile says so instead of showing one.
Improving · Safe to scale
support-bot
Error rate and max-token cutoffs both fell against the previous version, and the context loop closed with a fix Costile verified against the runs that followed it.
Degrading · Needs attention
claims-intake
Error rate and cutoffs rose against the previous version and the last fix has not been verified, so Costile keeps this agent out of the safe-to-scale count until the trend turns.
Start with proof, then scale control
Costile is for teams already running agents on live traffic, not for prototypes. Route a few production agents through it, see the first issues, and decide from evidence before adding broader governance.
What you get today
- Every incident arrives with the cause named, the evidence behind it, and a recommended fix
- Every agent carries a behaviour record across loops, retries, caps, and spend
- Every fix is verified against the runs that followed it
How it works
From trial to live traffic in three steps
Pick Individual or Team based on how many people need access, sign in with email and new password to create a workspace, and start the 10-day free trial without a card.
Send AI calls through Costile, review incidents and agent behavior, and only proceed to billing once the pilot proves the value.
Add payment details in Billing after the trial, manage your plan in Settings, and switch or cancel it at any time as team needs evolve.
Step two is one line of configuration. Costile mirrors the provider routes, so your existing calls keep working unchanged.
Setup
Anthropic and OpenAI SDKs
Replace - base_url = "https://api.anthropic.com"with + base_url = "https://api.costile.com"
Individual
For solo builders and operators improving production agents.
$79
/month
- Live incident diagnostics (cause, evidence, fix)
- Behavior visibility across agents and workflows
- Guardrails and caps, opt-in per agent
- Agent issue and incident alerts via email
Team
For teams needing governance, attribution, and shared operational evidence.
$299
/month
- Everything in Individual, plus:
- Team-wide attribution (person, API key, workflow)
- Audit-ready logs for compliance & procurement
- Fast alerts via Slack or webhooks
Enterprise
For organisations running agents across several teams, with procurement and security review.
Talk to us
- Everything in Team, plus:
- Unlimited API keys and members across workspaces
- Custom retention and data-handling review
- Security and procurement support, including the DPA
Good To Know
How does it work?
Route API traffic through Costile's proxy → it analyzes agent behavior in real time → returns cause + impact + fix reports.
Does routing through Costile add latency?
1.5 ms at p50, 2.5 ms at p99; 2.3 ms to first byte when streamed. Measured, not estimated —
node scripts/benchmark-latency.js reproduces it on your own hardware.
Do you store my prompts or responses?
Default: No. Only operational metadata is stored. Optional prompt capture can be enabled for debugging.
Can we inspect or run Costile ourselves?
Yes. The core runs locally and is fully inspectable. Cloud adds alerts, attribution, and managed ops.
What stage is Costile at?
Costile has a working proxy, dashboard, demo incident evidence,
authentication, billing, alerts, and audit-oriented pages. The
current work is pilot validation: proving the strongest workflows
and turning real customer feedback into a tighter product.
Why does this matter for Danish and European companies?
European teams adopting AI agents need operational proof, not just
model access. Costile is built from Denmark around traceable agent
behavior, metadata-first privacy, and evidence that supports
reliability, governance, and EU AI Act readiness.
Danish-built evidence for AI operations: Costile supports Article 9 risk-management and Article 12 record-keeping evidence for teams in scope, with automatic audit logs, request attribution, and root-cause capture for every incident.
Read the EU AI Act compliance page →