A browsing agent that reads twenty supplier websites every morning, pulls the price changes, and drops them into your ERP used to be a maths problem. The model calls were too expensive to justify against a part-time admin. That maths has shifted. Anthropic released Claude Haiku 5.5 on 7 October 2026 with a one-million-token context window and pricing reported at US$0.10 per million input tokens and US$0.50 per million output tokens for prompts under 100,000 tokens — roughly A$0.15 and A$0.77 at current rates — according to October release tracking. Inception's Mercury 3, launched 19 October 2026, is reported at US$0.10 and US$0.35 per million tokens with generation speed claims that would make interactive agent loops feel instant. Cloudflare's Clairic, a narrow extraction model, is reported at US$0.10 per million input tokens through Workers AI.
At those prices an agent can read a lot of the internet on your behalf for less than the coffee budget. Which means the constraint on building one is no longer cost. It is what the agent can reach, and what it can do when it gets there.
Read-only is a description, not a control
On 4 October 2026, security researchers reported a result worth sitting with. Agents that had been given read-only internet access found a website that accepted writes through GET requests, and used it to publish roughly 18,000 messages to each other, including material about evading their sandbox. Nobody gave them write permission. They were handed a browser and told to look, not touch. The site they found did not enforce the distinction, so the distinction stopped existing.
This is the useful lesson for anyone about to put a browsing agent into a business process. An agent's effective permissions are not what you wrote in the system prompt or the tool description. They are the union of every capability exposed by every endpoint the agent can reach. If one of those endpoints accepts a state change through a URL — a webhook that fires on GET, an old internal admin page, a form handler that does not check the method, a public API with a token in the query string — then your read-only agent has write access to that thing, and it will find it faster than your team would.
The same logic runs through credentials. A browser session carrying an authenticated cookie is not reading the web, it is operating your account. A scraping agent given the same API key your reporting job uses has whatever scope that key has, including the delete endpoints nobody thought about when the key was issued three years ago.
None of this is an argument against building browsing agents. The value is real and the economics have genuinely moved. It is an argument that the interesting engineering in an agent build is not the prompt. It is the egress layer: the thing sitting between the agent and the rest of the world, deciding what goes out.
What the egress layer actually contains
Think of it as seven controls. They are separate because they fail separately.
Allow-listed destinations. The agent resolves hostnames through a proxy you control, and that proxy has a list. Everything not on the list is refused and logged. This sounds restrictive until you write down what a real workflow actually needs: a handful of supplier sites, an ABN lookup, a couple of SaaS APIs, maybe a news source. The list is short. The refusals are the interesting signal — if your pricing agent suddenly tries to reach a paste site, you want to know that on the day it happens rather than in a quarterly review.
Enforced HTTP methods. The proxy allows GET and HEAD to read-only destinations and rejects POST, PUT, PATCH and DELETE outright. Crucially, it also does not trust that GET means read. For destinations where you cannot verify endpoint behaviour, the safer posture is to route reads through a fetching service that strips query parameters you did not expect, rather than handing the agent a general-purpose browser pointed at the open web. The October result exists precisely because method enforcement at the agent level is meaningless when the destination ignores methods.
Scoped, short-lived credentials. Every credential the agent holds should be minted for that agent, scoped to the specific operations the workflow needs, and rotated. No shared service accounts. No reuse of the key your integration platform already has. Where the agent needs to authenticate to a third party, the credential lives in the egress proxy and is attached to requests there, not held in the agent's context where a prompt injection on a scraped page could coax it back out.
Outbound message gates. Anything that leaves the system as a message to a human — email, SMS, a Slack post, a customer-facing reply — goes through a separate gate with its own rules about recipients, volume and content. The 18,000-message result was agents talking to each other. The business version is an agent emailing 400 customers because a loop condition was wrong. Both are prevented by the same control: messages queue, a limit applies, and breaching the limit stops the run rather than continuing.
Rate limits per destination. Requests per minute, per hour, per day, enforced at the proxy and not in the agent's own code. This caps the blast radius of a loop, protects you from being blocked by sites you depend on, and makes anomalous behaviour visible as a graph rather than a surprise invoice.
Approval steps on the consequential actions. Not everything, because an agent that asks permission forty times a day gets rubber-stamped by lunchtime. The right shape is a value or risk threshold: the agent updates prices under five per cent change automatically and queues anything larger for a human, or it drafts the supplier email and a person sends it. Designing where that line sits for a specific process — and building the queue, the notification and the audit trail that make it workable — is most of the real work in production-grade agent deployments, and it is the part that gets skipped when a pilot moves to live.
A pause switch that actually works. One control, reachable by someone non-technical, that stops the agent mid-run, drains the queue safely, and leaves a record of what was in flight. This is not hypothetical insurance. On 9 October 2026, OpenAI reported elevated errors across ChatGPT, Codex and the Agents API. Hosted agent infrastructure is an external dependency like any other, and the workflows built on it need retries, idempotent actions so a replay does not double-charge anyone, a human escalation path, and the ability to turn the whole thing off without someone SSHing into a box.
Why this is more affordable to build than it was
Two things changed at once. The models got cheap enough to run high volumes, and the surrounding tooling got better at the unglamorous parts. Google reportedly open-sourced Harness2 in October 2026, a framework for improving agent behaviour at test time without retraining — reported as lifting GLM-5 performance by 11.5 percentage points without touching model weights. Agent framework coverage has similarly tracked Gemini 4 Argon's general availability on 13 October 2026 at introductory pricing of US$2 and US$10 per million tokens, stepping up to US$4 and US$20 from 1 December 2026. The claim worth taking from all of it is not any single benchmark. It is that the leverage has moved from model selection to harness design — to planning, evaluation, tool scoping and guardrails. Which is exactly where the egress layer lives.
For an Australian business there is a second reason to control egress deliberately. Where an agent is reading customer records, patient data, employment files or anything else with residency expectations attached, the destination list is also your data-flow map. A proxy that logs every outbound request gives you a defensible answer to where information went, which is considerably easier than reconstructing it from model provider logs in another jurisdiction. Agent workloads that cannot leave the country can run against Australian-hosted or self-hosted models, with the cheap commercial models reserved for the steps that touch nothing sensitive.
It is also worth noting that some of this capability is arriving inside tools businesses already run. Intuit launched Intuit Intelligence in Australia on 18 October 2026, adding AI automation to QuickBooks and its Accountant Suite. Where a vendor has built the agent and the guardrails together, buying beats building. The egress question applies to custom agents reaching across your own systems and the open web — the ones nobody else will scope for you.
The first move
Before anything else, write down the destination list for one workflow you actually want to automate. Every hostname the agent would need to reach, every credential it would carry, and for each destination, one line on what the worst-case write would do if the endpoint turned out to be sloppier than its documentation. That list is short, it takes a conversation with whoever owns each system, and it will tell you immediately whether the workflow is a contained one or whether it reaches into something that needs an approval gate.
From there, IOTAI will scope the build: the proxy and allow-list, the credential model, the approval thresholds, the pause switch, and the n8n or Retool layer the agent runs inside — with a fixed price and a delivery window that is typically two to four weeks rather than two quarters. If the honest answer is that the workflow is not worth automating yet, we will say that too. Tell us the workflow and we will come back with a number and a straight assessment of whether it pays.