Australian firm opens the books on five months of AI agent controls
After five months on its agent tronic, Agentic Digital Identity has detailed receipts, a nine-check send gate and permission tiers — plus the silent failures it caught, including 86% of tokens burned by a default heartbeat.

Why it matters
- ADI detailed controls for its internal AI agent tronic after more than five months of daily operations.
- Every outbound message passes a single send gate running nine checks, and no message ships without recorded human approval.
- The first write census missed seven files; one could create, change and delete calendar events with no record kept.
- Over one seven-day period, 86% of fresh model tokens were consumed by an unconfigured default background heartbeat.
- A write ledger has been checked against external providers four times a day since August.
Five months of running day-to-day operations on an AI agent has produced something rare in the industry: a public account of what actually broke. Agentic Digital Identity (ADI), an Australian firm, has outlined the controls it uses to govern its own internal agent, tronic. The same controls ship inside the agents it builds for clients.
The details landed on 18 November 2026. They cover a small company eating its own cooking, and the failures it found are the kind most firms never disclose. Founder and engineer Thomas Beard said the main effort since launch has focused not on the language model itself, but on the mechanisms around it that determine what the system can do and how its actions get verified.
The stakes reach beyond one Australian firm. Businesses everywhere face growing pressure to put controls around autonomous software agents, and Beard pointed to a gap between what large companies can buy and what smaller firms can access when they deploy their first agent.
"A large enterprise can buy a governance layer," Beard said. "A smaller business running its first agent, however, needs those controls built into the agent from day one."
What does the agent actually do?
The internal agent, tronic, handles routine office work. It drafts outbound email, turns meeting transcripts into minutes and action items, runs approved tasks from a project board, and reports into Slack. That's a modest footprint. But every one of those tasks touches a system of record, and each touch is a chance to silently change something.
Beard's framing is blunt: code does the doing, and the model does the talking.
"Plain code does the work and the model handles the language," Beard said. "A job only counts as done when there's evidence it happened."
How do you prove an agent succeeded?
ADI's central rule is that a task counts as successful only when there is deterministic evidence of the result. That means either a process exits cleanly and creates the expected output, or an external system returns a real record ID. Confident model claims don't count.
Since August, the agent's main write paths have returned a receipt stored in a ledger. That ledger is then checked against the external provider four times a day.
"Models will report success, pretty confidently, when they've failed," Beard said. "That's why a write now either produces a receipt or we treat it as not having happened."
This is the quiet heart of the story. LLM agents are probabilistic narrators of their own performance, and the industry's default posture has been to trust the narration. ADI's answer is to make the receipt, not the report, the unit of truth.
What sits in front of every outbound message?
Every outbound message passes through what ADI calls a single send gate. The gate runs nine checks covering:
- approval
- approver
- payload
- mode switch
- send whitelist
- recipient policy
- budget
- suppression list
- transport
No message goes out unless a person has recorded approval. Approved content is fingerprinted and checked again before release, so it cannot change after review. The design closes a classic gap: content approved once, then quietly modified before it leaves the building.
What is a write census, and why did it fail?
Another control ADI described is a write census: a list of every point in the code where the agent can alter something in an external system. The list is generated by scanning the codebase, and a test fails if the list and the code do not match.
That process exposed gaps in an earlier version of the census itself. Beard said the first census missed seven files because it recognised only one of two common code shapes.
One of those seven files could create, change and delete calendar events without any record being kept. It's a small hole with real consequences, and it shows the low-visibility risk that emerges when agents plug into everyday business tools.
"Nobody would have noticed until a meeting went missing," Beard said. "You can't log only the important ones and call it an audit trail."
What failures looked like success?
The company's work on operational controls also surfaced problems that masqueraded as successful runs. In one case, 17 scheduled jobs were configured so that a crash counted as a successful exit. The failure alert never triggered because, as far as the system could tell, nothing had failed.
ADI also found an inbox watcher that used Gmail's unread flag to determine what it had already seen. If a reply was first read on a phone, the message disappeared from the watcher's view and was never processed again.
Then came the cost discovery. Over one seven-day period, 86% of fresh model tokens were consumed by a background heartbeat that was enabled by default in the runtime and left unconfigured. Nearly nine-tenths of the agent's token spend was going to nothing user-visible.
Beard said the discovery forced a stricter default approach: any missing setting that could affect a person-facing action is now turned off.
"Now if a setting's missing and it could reach a person, it's off," Beard said.
How does the agent earn new permissions?
The agent's permissions are tied to a scoring system based on its own run log. Clean runs with evidence earn points. Failures and incidents reduce the score. Once the record qualifies, the scoring system can flag the agent as eligible for a new permission tier.
The human stays in the loop where it matters. A person must still approve any move to a higher tier, and only a human can make that change. The design resembles a trust ladder: the agent earns the right to ask for more rope, but never grabs it itself.
ADI said it applies the same structure to client work, including the approval gate, the evidence-based success tests and the write ledger. Each client's names, figures and operating rules remain separated within that client's deployment.
Why does this matter beyond one firm?
ADI works with businesses in Australia and the United Kingdom, installing and maintaining AI agents inside existing workplace tools. That position gives Beard a view of where the market's anxiety has moved. The question, he said, is no longer whether a model can generate text. It's whether the organisation can trace what the software was allowed to do and what it changed.
"Look, the model is the easy part now," Beard said. "What a business actually needs to know is what the agent's allowed to do and who said yes, and then what it changed."
The company's receipts, send gates, write census and permission tiers are one small firm's answer. But the failures it disclosed — silent calendar writes, 17 jobs failing quietly, 86% of tokens burned on a default heartbeat — are the concrete case for why such controls need to be built into agents from day one, not bolted on after the first meeting goes missing.
Source: itbrief.com.au
More from Elena Vasquez
Show full bio
Market editor covering technology and product news in artificial intelligence and machine learning at AI In Context.
708 articles
Related articles
- Australia Leads AI Agent Adoption at 46%, but Managers Pay the Price
- 83% of Firms Run Unapproved AI Coding Agents, BlueFlag Research Finds
- Accenture and OpenAI Team Up to Put Agentic AI at the Core of Enterprise Operations
- Appian Claims 80% Token Cuts by Putting AI Agents Inside Processes
- The EU's Next Chapter in AI Is Agentic, New Analysis Claims