All insights
10 Apr 2026·11 min read·By Rajneesh Malviya
Agentic systems without the handwaving
Most "agent" architectures in 2026 are still LLM-loop-with-tools dressed up. Here is the framing we use to decide when an agent is justified, and the four traps that show up the moment you put one in front of real users.
Agentic systemsArchitecture
An agent is justified when the problem has open-ended state — the system needs to decide what to do next, multiple times, against a partially-observable world. Everything else is a workflow.
The four traps
- Latency cliff: the moment your tool count crosses ~12, the planning step gets slow and the user notices.
- Reasoning loops: the model decides to "verify" itself indefinitely. Hard cap on iterations is the only thing that has held up for us.
- Tool fan-out: the agent calls three tools in parallel that each take 4 seconds, and the user is staring at a spinner. Front-load the user with a streaming "I'm checking three things" rather than silence.
- Memory rot: long sessions accumulate stale state. Resetting at semantic boundaries (new task, new doc) beats letting it grow forever.
We've shipped four agentic systems into regulated buyers. The architecture diagrams all look like the public ones. The thing that makes them work is the boring fence-posts around the loop — caps, resets, tool-call budgets, and a really good "abort" path.