TryAgent

The platform

Everything the autonomy lifecycle needs.

TryAgent discovers every risky call, guards it at the boundary, decides against a versioned rubric, routes escalations to the right human, graduates what earns trust — and proves all of it. Every capability is included on every plan.

Start freeScan your repo — no signup

Discover

Find every risky call before it finds you

tryagent scan reads your repo and maps the agent surface — LLM calls, tools, side effects — deterministically, with no API key.

  • Deterministic detectors. Unguarded side effects, unvalidated LLM outputs, and rubric-less calls — resolved across file imports, with exit codes for CI.
  • A wizard that drafts, a human who approves. scan -i drafts each rubric — intent, examples, conditions, output schema — and wires guard() as a reviewable diff.
  • Governance in code review. tryagent review and the GitHub Action block PRs that remove approvals or add unguarded side effects.
Run your first scan
terminalno api key

npx @tryagent/cli scan

[HIGH]unguarded side effect · stripe.refunds.createtools/refunds.ts

[MED]unvalidated LLM output · claim-classifieragents/claims.ts

[MED]missing rubric · payments.refundtools/refunds.ts

3 findings · run scan -i to draft rubrics and wire guards

Guard

One line in code — or a config file

guard() wraps any tool and pauses it at the boundary until the decision lands. The MCP proxy fronts any MCP server. Adapters cover LangGraph, the Vercel AI SDK, and OpenAI Agents.

  • One line. tryagent.guard("payments.refund", refundTool) — policies and rubrics auto-register on the first call, so there is no dashboard prework.
  • Zero code changes over MCP. Map upstream tools to policies in one config file; unmapped tools pass through untouched.
  • Fail closed. Deny is the default on SLA breach, and if TryAgent is unreachable, guarded tools don't execute.
Read the quickstart
tools.ts
export const tools = {  refund: tryagent.guard("payments.refund", refundTool),};

or over MCP — no code changes

"policies": { "stripe:create_refund": "payments.refund" }

Decide

A versioned rubric on every decision point

Each LLM call gets a rubric: the contract for what it may do, with deterministic conditions that decide at runtime whether it proceeds, gets reviewed after, or needs a human first.

  • Deterministic conditions. "Autonomous below $100, human required above" is plain math over structured evidence — no AI in the hot path, testable in CI.
  • Shadow mode. Turn TryAgent on in production with zero behavior change: every call executes, everything lands in the review queue.
  • No deploys. Conditions, presentation, and modes are rubric versions — edit on the web, roll back by reactivating the old version.
How rubrics work
payments.refundrubric v3
ifargs.amount > $500
required
ifargs.amount > $100
reviewed
ifelse
autonomous

Missing or unparseable evidence resolves to the most restrictive mode on the version — ambiguity keeps the human in the loop.

Route

The right human, with the full picture

Policies own the route: queue, reviewers, on-call chains, SLAs, and timeout defaults — versioned and auditable, never wired into agent code.

  • Reviewers and on-call. Tier reviewers and define escalation chains so the right person always sees it — in the inbox, Slack, or your own surfaces.
  • SLAs and defaults. Deadlines and a default choice so a run is never stuck waiting on a human.
  • Precedent rides along. A recommendation learned from your resolved escalations, with confidence in plain view — scoped to your tenant, never shared.
How routing works
refunds.high_valuepolicy v3
queueFinance on-callSLA 2h
chainowner at 0m → admin at 15mauto-page
deliverinbox · slack #finance-approvals
timeoutdefault choice: rejectfail closed
recommendationapprove · 86% confidence

4 of 5 similar refunds approved by finance

Graduate

From required review to earned autonomy

The review queue is designed to shrink: four modes on the rubric — required, reviewed, sampled, autonomous. Deterministic criteria over real decision history propose the promotion; a named human approves it; demotion is automatic.

  • Threshold math, not model vibes. A rolling window, a minimum decision count, and the override rate over reviewed decisions decide readiness.
  • Promotion is an escalation. "0 overrides in 214 decisions — graduate to sampled?" arrives with the stats as evidence and a named approver.
  • Automatic demotion. A severe override or an override rate above your bar drops the rubric a rung and pages the owner.
The graduation model
payments.refund · rubric v4ready to graduate
  1. required
  2. reviewedcurrent
  3. sampled 10%
  4. autonomous
214 decisions0 overrides30d window

Prove

An audit trail that survives the auditor

Every decision, review, and mode change writes to a tamper-evident, per-tenant hash chain. You don't just log everything — you can prove nothing was edited.

  • Verifiable end to end. Each audit event hashes the one before it; one call recomputes every link in the chain.
  • Exports on a stable schema. Decision logs as CSV or NDJSON with the full event chain — sync for quick pulls, async jobs for large ranges.
  • SIEM and retention. HMAC-signed streaming to your SIEM, per-tenant retention windows, and a legal hold that blocks deletion.
Security overview
audit chain · esc_4821
createdrun_7f3 · payments.refund8c41…
auto_decidedrubric:reviewed · approve2fb9…
reviewedagree · j.leed07a…
exportedNDJSON · 30d range · admin5e2c…
GET /audit/verify → chain intact · 48,112 events

Put a rubric on your riskiest call today.

100 escalations free each month, every capability included, no credit card. Upgrade when your agents create more decision load.

Start freeCompare plans