Skip to content
in/guard/out
in/guard/out

Agent evaluation & observability

in/guard/out vs Maxim AI

Maxim AI is a serious evaluation platform: agent simulation across personas, machine and human-in-the-loop evals, multi-turn session observability, alerting. All of it shares one property – it measures. in/guard/out’s job starts where measurement ends: standing in line with production traffic and deciding, per request and per tool call, whether this response ships, this payment executes, this run continues. A trace of the incident is not a block of the incident.

Side by side

CapabilityMaxim AIin/guard/out
What it isSimulation, evals, and observability for agentsA runtime control plane that enforces on live traffic
When it actsBefore release (simulation/evals) and after the fact (monitoring, alerts)In line – before the response reaches the user or the tool call executes
A hallucinated figureScored in evals; alerted on in productionCaught deterministically and repaired or blocked on the wire
An ungrounded paymentVisible in the trace afterwardsBlocked before execution – the payee must appear in the run’s sources
A runaway agent loopAlert fires while it burnsBudget gate stops the next call – cost capped mid-run, atomically
PII to the providerObserved in tracesScreened to placeholders on the way in, restored on the way back
Tool-call policyTraced and evaluatedEnforced: schemas, permission tiers, ordering, taint – per key, per run
Audit outputDashboards and tracesA run graph of enforced decisions – what was allowed, repaired, blocked, and why

Enforcement and evaluation compose

This is the least adversarial page on this site: teams should do both. Evals catch what you can reproduce before you ship; the wire catches what production does anyway. The two feed each other – in/guard/out ships its own offline eval harness, and every violation it catches in production is a labeled, replayable case for your evaluation suite, whichever platform runs it. (Maxim also builds the Bifrost gateway – that lane is covered in vs Bifrost.)

Frequently asked questions

Does in/guard/out replace an eval platform?

No. Evals measure quality before you ship and trend it after; in/guard/out enforces policy on live traffic. Run both: the proxy’s production catches – each one a real failure with full context – make the best eval cases you will ever write.

Maxim observes multi-turn sessions – isn’t that the run graph?

It’s the observability half of it. A session trace shows what happened; in/guard/out’s run graph shows what happened and what was enforced – the budget that stopped step 14, the payment that was blocked, the number that was repaired – because the same layer that drew the graph made the decisions.

Can they run together?

Cleanly. in/guard/out sits on the wire; Maxim’s SDK instruments your app. Neither interferes with the other, and the proxy’s per-stage findings can be exported for analysis alongside your traces.

Evals find it in testing. The wire stops it in production.

in/guard/out enforces in line while your eval platform measures - and every production catch makes a new test case. We are running a limited demo - sign up and we will get you in as soon as we can.