Give Every Agent a Stop Condition and a Recovery Path
Define success, lack of progress, cancellation and escalation before running an agent, so repeated activity does not become an endless substitute for completion.
Agentic Systems 3 min read
Agentic Systems
Tool-using agents and AI coding workflows: server-side authorization, budgets, and the boundaries that keep an autonomous action safe.
13 posts
Conviro Agent Control Room: Making Agent Work Inspectable
Why I built an internal Agent Control Room with activity, findings, spend visibility and explicit human approval boundaries.
Agentic Systems 4 min read
Separate an Agent’s Recommendation from Its Side Effects
Keep proposed actions, authorization, execution and confirmed outcomes distinct so an agent can explain what it intends to do and what actually happened.
Agentic Systems 3 min read
AI Coding Agents Need a Verifiable Deployment Boundary
Connect the code reviewed, the code tested, and the artifact deployed so an AI-assisted workflow cannot confuse completed work with a safe release.
Agentic Systems 3 min read
Build Research Agents That Preserve Their Evidence
A research agent should preserve where each claim came from, when it was observed and what remains uncertain as findings become summaries.
Agentic Systems 3 min read
Delegate to Multiple Agents Only When the Boundary Is Clear
Useful multi-agent work starts with independent tasks, bounded authority and an explicit merge decision, rather than adding more roles to the same problem.
Agentic Systems 3 min read
A Tool-Using Agent Needs Server-Side Authorization
Treat model-generated tool calls as proposed actions. Validate permissions, scope, arguments, budgets, and approval at execution time.
Agentic Systems 3 min read
Human Approval Should Expire When the Plan Changes
Tie approval to a concrete action, target, input version and lifetime, so an agent cannot carry permission from one proposal into a different operation.
Agentic Systems 3 min read
Make Long-Running Agents Resume from Durable State
Design agent jobs around persisted decisions and accepted effects, so an interruption does not force the system to guess what already happened.
Agentic Systems 3 min read
Treat Retrieved Instructions as Untrusted Data
Keep instructions found in documents, websites and tool results from becoming authority over an agent’s task, tools or destination choices.
Agentic Systems 3 min read
Agent Memory Needs Expiry, Provenance and Scope
Treat agent memory as scoped evidence with a source and lifetime, so remembered preferences do not silently become current facts or permissions.
Agentic Systems 3 min read
Write Tool Contracts That Survive Ambiguous Requests
Reliable agent tools need clear meanings for identity, filters, missing fields, result completeness and errors, beyond valid structured input.
Agentic Systems 3 min read