Belal Shafiq

building AgentTrafficCop. the runtime interceptor for AI agents. catches infinite loops, validates tool calls, and stops your API budget from burning. live now.

[Belal Shafiq @bshafiq_]

agents don't fail because they're dumb. they fail because they loop until your api bill hits $500.

built AgentTrafficCop to sit on the spend path and kill the execution after 3 identical calls.

live today. link in bio.

[Belal Shafiq @bshafiq_]

@lostinazeroy yeah, 422 saves the orchestrator from a total crash.

lmk if you ever want to ditch the custom logic and test ours. good luck with the builds

[Belal Shafiq @bshafiq_]

@alex09x_dev totally fair. a timeout just kills the workflow, it doesn't resume it. sounds like you need a state machine or durable execution engine more than just a loop breaker. good luck with the setup man.

[Belal Shafiq @bshafiq_]

@NicoHao4s @ZeeshhDecodes evals are good, but they run after the fact. for payments, you need a runtime check before execution.

are you hashing the payload at the handoff to ensure it doesn't double-fire?

[Belal Shafiq @bshafiq_]

@zemnanet the 3am hang is brutal. it just sits there burning resources while you sleep.

you need a breaker that throws a 422 if it hangs, instead of waiting for a human that isn't there.

where are you logging these denials right now?

[Belal Shafiq @bshafiq_]

@imehr yep. rigid middleware is the bottleneck. a blind retry limit can't tell a bad prompt from an A->B->A loop.

you have to hash the payload state at the handoff layer to actually adapt.

how are you validating the outer loop right now?

[Belal Shafiq @bshafiq_]

right now, an agent loop costs you a $3k api bill.

by 2028, a fortune 500 will run 150,000 agents.

a loop then won't just burn tokens. it'll execute duplicate refunds, trigger bad trades, and send bad emails.

build the breaker before you build the swarm.

[Belal Shafiq @bshafiq_]

@SalahMakboul06 the "in doubt" state handles crashes perfectly. but does it catch conversational loops?

A asks, B denies, A asks again. no crash, just A->B->A until the api budget dies. are you hashing the payloads to catch that, or just logging the calls?

[Belal Shafiq @bshafiq_]

@Vinayak1616 point 2 is the key. putting guardrails in the prompt is just asking the llm to police itself.

putting a deterministic check outside the graph to block duplicate refunds or state loops is the only way to actually make it safe. how are you handling the validation layer?

[Belal Shafiq @bshafiq_]

@olegzanx @_BananaCrystal schema drift causes a validation fail, which causes a retry, which turns into a silent loop. are you building a strict validation layer at the connector for this?

[Belal Shafiq @bshafiq_]

@harden_369741 @bebuu_ai exactly. a louder alarm at 3am is still an alarm at 3am. catching it at step 3 means the alarm never goes off. sleep > everything lol

[Belal Shafiq @bshafiq_]

@harden_369741 @bebuu_ai haha "negotiating with it" is too real.

the whole goal is to kill the loop before the first alert so you can just roll over and go back to sleep. no negotiating required.

[Belal Shafiq @bshafiq_]

@BoboLinNotes orchestrator pattern is the right move. but A->B->A handoff loops are brutal.

to the orchestrator, it just looks like a normal retry. traces look fine, but your api budget is draining fast.

how are you tracking state across the agents to catch the cycles?

[Belal Shafiq @bshafiq_]

@harden_369741 @bebuu_ai lol zombie code is undefeated. see you at 3am

[Belal Shafiq @bshafiq_]

@harden_369741 @bebuu_ai haha honestly fair. we've all got that one janky script we refuse to let go of.

just know the door's open when it inevitably blows up on a 3am run lol.

[Belal Shafiq @bshafiq_]

@harden_369741 @bebuu_ai respect the ugly tracker, it saved your wallet.

we actually built a middleware that does exactly that, so you don't have to maintain custom state trackers for every graph. it's in my bio if you ever want to ditch the custom code and test it on your next run.

[Belal Shafiq @bshafiq_]

@sanathbhat39 @anah_sahh wiring it into the runner loop is the right move. never let the agent police itself.

we actually packaged that exact ring buffer logic into a middleware so you don't have to maintain it manually. link is in my bio if you ever want to test it or try to break it.

[Belal Shafiq @bshafiq_]

@sanathbhat39 @anah_sahh "agents lie about their own state" is the best way i've heard it put.

if it can bypass the check, it will. the ring buffer has to live outside the graph.