agent action proof

Heso Witness

Policy, approval, and proof for every web action an agent takes.

Let agents read, fill, and submit across regulated external systems — payer portals, banking, insurance, procurement, vendor and SaaS admin panels — with every action checked against policy, approved when needed, signed, and replayable.

Start with one workflow: claims, eligibility, payments, exports, or any submit that crosses a compliance boundary.

live execution receipt
claims-agent-v2 / run_4829
policy: enforced
01
open
payerportal.com/member/eligibility
02
extract
member status, plan, deductible
03
fill
claim amount, CPT, diagnosis fields
04
submit
claim filing button
05
receipt
ed25519:9d41 / replay ready
observed hash
b3:81f9
final hash
b3:44aa
replay
available
Policy Gateway

Stop the action before it lands. Domain, field, spend, and submit rules enforced before the agent can do anything irreversible.

Approval Queue

Keep a human in the loop where it matters. Claim filing, payments, exports, and account changes route to a reviewer before they happen.

Signed Receipts

Hand audit an artifact, not a story. Each action carries URL, agent, model, DOM state, policy decision, approver, and final state — verifiable offline.

Replay

When something breaks, rebuild the run from captured state. No model memory, no guesses — just the exact bytes the agent saw.

where it sits

Between your agent and the external web.

Witness is the security boundary that sits between your agent and anything your security, compliance, or legal teams would want to see before it happens — and have a defensible record of after.

01

AI agent

task, identity, model, account

02

Heso Witness

policy check, approval gate, receipt envelope

03

Heso runtime

native execution, cassette, hashes, browser fallback when needed

04

External portal

payer, vendor, SaaS, bank, internal app

05

Audit systems

SIEM, Datadog, Splunk, legal export, security review

production exposure model

Count the actions security has to trust.

The question is not browser minutes. It is how many external web actions need proof, approval, replay, and a defensible audit trail before agents can run in production.

production exposure model

External portal actions under control

payer portal workflow
sensitive data
68%
side effects
4.5%
failure replay
1.8%
checks / run
5

This models production exposure, not labor replacement: how many external web actions need policy, proof, approval, and replay.

monthly control surface
signed receipts
12K
one per portal run
approval gates
540
submit, payment, export, account change
sensitive-data receipts
8.2K
customer / regulated data touched
replayable failures
216
failed runs with reconstruction material
annual policy load720K checks / year

The budget case is not browser cost. It is the ability to let agents touch production portals because security can see the rule, the approver, the signed receipt, and the replay path.

control plane modules

Everything around the action.

01

Policy Gateway

Define what agents can touch before they touch it. Domain allowlists, field-level rules, spend limits, and submit gates — written once, enforced on every run.

02

Approval Queue

Send the high-impact actions to a human first: claim filing, payments, purchases, data exports, account changes, delete operations, final submits.

03

Signed Receipts

Cryptographic proof of what the agent saw, what it decided, who approved it, and how it ended. Verifiable offline, defensible in audit.

04

Replay & Incidents

When something goes wrong, rebuild the exact run from captured state — no model memory, no second-guessing. Forensics in minutes, not days.

05

Enterprise Controls

SSO/SAML, SCIM, RBAC, redaction, credential custody, customer-managed keys, audit retention, SIEM export, and private/VPC deployment.

06

Operator Dashboard

One pane of glass across every agent action: filter by team, account, model, workflow, policy decision, approval state, or failure reason.

why not the existing stack?

Browser infra and traces are necessary. They are not the proof layer.

Witness adds
pre-action policyyes
signed receipted25519
state integrityblake3
offline replaycassette
positioning
  • Browserbase gives agents real browser sessions, identity, logs, and replay. Witness is the policy and proof layer before high-impact web actions.
  • LangSmith, Datadog, and Langfuse show agent traces. Witness records whether an external web mutation was allowed, blocked, approved, and signed.
  • Generic audit logs say an event happened. Witness receipts can be verified offline and tied to state hashes and replay material.
  • Playwright traces help developers debug. Witness is built for security, legal, compliance, and operations when agents touch third-party systems.
enterprise evaluation

Prove one workflow end to end.

Bring one workflow where agents touch real accounts, customer data, forms, or submit buttons. The evaluation proves controls and evidence on a single workflow before you scale it across more.

one workflow

Bring a real workflow where agents touch external accounts, customer data, forms, or submit buttons — claims, eligibility, payments, exports, account changes.

policy file

Define allowed domains, field rules, submit gates, export rules, and approval requirements.

proof check

Verify a signed receipt offline and replay one failed or blocked run from captured state.

security export

Confirm SSO/RBAC shape, redaction needs, SIEM export, retention, and private deployment.

evaluation criteria

A policy blocks or routes submit before impact, a signed receipt verifies offline, a failed run replays from captured state, and security can export the evidence trail.

schedule a workflow review