Agents · Egress · Evidence

Agents need action controls and traffic evidence.

Microsoft AGT is the right reference point for deep in-process action governance. Talon’s role is different: put an EU-sovereign evidence gateway around the agent’s LLM and tool traffic so you can prove what data left, where it went, and which policy allowed it.

Run the 60-second demo Compare vs AGT Evidence model

Agent framework
  → proposes prompt + tools
  → Talon gateway / MCP proxy
      → scan PII
      → check sovereignty
      → enforce budget
      → filter or deny tools
      → record decision
      → sign evidence
  → provider or tool

AGT: in-process actionsTalon: egress evidenceUse both when neededEU reviews

Layer model

Do not confuse agent-action governance with traffic governance.

AI agents add risk because they can reason, call tools, and run loops. Talon does not claim to replace deep in-process action governance. Talon covers the evidence-grade egress layer: LLM calls, provider routing, PII/content policy, cost, tool exposure, and exportable proof.

🏗️

Action layer

Use AGT-style controls when you need deterministic interception inside the agent runtime.

🛡️

Egress layer

Use Talon when you need a chokepoint for model/provider traffic and third-party tool access.

🧾

Evidence layer

Use Talon when the buyer asks for signed records, auditor exports, and integrity verification.

Agent risks

What Talon proves around agent workflows.

Risk	Example	Talon control
PII leakage	Support ticket contains IBAN or email.	Input/output scan, redact/block/warn policy, signed evidence.
Non-EU egress	Confidential prompt sent to a non-approved provider.	Provider metadata, EU strict/preferred/global modes, routing evidence.
Runaway spend	Agent loop calls the model repeatedly.	Pre-spend budget checks and per-run / per-caller limits.
Tool exposure	Model sees `delete_records` or `export_all_customers`.	Allowed/forbidden tool policy, filtering/denial, evidence record.
No proof after review	Only app logs exist.	HMAC-signed records, hashes, CSV/JSON/signed export.

Best use

Use Talon around existing agents first.

Talon’s near-term wedge is strongest when an EU team already has agent or LLM traffic and needs to pass a DPO, customer, security, or audit review without rewriting every agent.

First workflow to govern

- support bot
- Slack bot
- coding agent
- LangGraph workflow
- OpenClaw setup
- vendor MCP integration

Goal: signed evidence from day one

Next step

Put Talon around one agent workflow.

Route its LLM calls through Talon, set one PII policy, set one cost cap, and verify the evidence.

Run the demo LangGraph integration