Agentic Security at Trent: From Judgment to Time-Bounded Delegation

DOAgents, consistent reasoning limits, and the power of ‘I don’t know’

Neil D. Lawrence

Trent AI Offsite

Context and objective

Building on our previous Trent session

  • Last time: explicit intent and shared context reduce misimplementation.
  • This time: extend that logic from single assistants to networks of agents.
  • Core question: how do we keep institutional judgement while scaling delegation?

Security at machine speed

  • Agentic systems turn language into actions (tools, APIs, workflows).
  • Teams ship faster; attack surface and ambiguity both expand.
  • The bottleneck shifts to judgement, coordination, and timely escalation.

Institutional tacit knowledge

  • The “judgement layer” in organisations is often tacit: norms, exceptions, escalation paths, and context that rarely makes it into documentation.
  • It lives in handoffs and approvals: what gets challenged, what gets waived, and what triggers a halt.
  • Ceding this tacit knowlege without making it explicit is how we accumulate agentic debt.

Architecture: DOAgents

DOAgents for agent networks

  • Use a data-oriented interface between agents: shared state, explicit contracts, typed handoffs.
  • Model work as subgraphs: retrieval, synthesis, planning, tool-use, verification.
  • Analyse at the graph level, not just individual prompts.

Why this helps Trent now

  • Makes the judgement layer inspectable: what each agent saw, decided, and handed off.
  • Enables selective autonomy: low-risk paths can run fast, high-risk paths route to review.
  • Creates a foundation for measurable SLOs on security outcomes, not just latency.

Reasoning limits and trust

Consistent Reasoning

  • Assume an agent \(R\) with two properties: Logical consistency \(R\) never believes both \(P\) and \(\not P\).

  • Trust in its own reasoning If \(R\) concludes something by valid reasoning, it believes that conclusion.

  • This models an ideal thinker or reasoning AI.

Consistent Reasoning Paradox (CRP)

  • The paradox shows that an agent that is:

  • logically consistent

  • fully reflective about its own reasoning

  • perfectly trusting of its conclusions

  • cannot maintain all those properties simultaneously.

The missing primitive: “I don’t know”

  • Agents need an explicit I don’t know action, not just low-confidence prose.
  • “I don’t know” must be operational: halt, escalate, or request additional evidence.
  • This is a control primitive for safety, not a model weakness.

Paying down agentic debt

Agentic Debt

  • Agentic AI could pay down technical and intellectual debt.
  • But it can create agentic debt: delegation without authority/authorship

Agentic Debt

  • Delegation of workflows without crisp boundaries.
  • Agentic debt is about unsafe or illegible delegation.

Time-bounded delegation in DOAgent graphs

  • Assign each node/subgraph a time budget \(\tau_i\) and termination policy.
  • At timeout: complete with evidence, or emit I don’t know and escalate to human.
  • Optional prompt augmentation: agents see remaining time and adapt search depth.

Choosing time budgets (\(\tau_i\))

  • Tune empirically from traces: success rate, escalation rate, and incident outcomes.
  • Optimize expected cost: human interruption cost vs compute waste vs risk penalty.
  • Different tasks need different \(\tau_i\): triage may be short; remediation planning longer.

20-minute takeaway

  • Institutional tacit knowledge is the judgement layer; don’t silently cede it.
  • DOAgent-style graphs make delegation explicit, inspectable, and governable.
  • “I don’t know” + time-bounded escalation can convert agentic debt into managed risk.

Thanks!

References