Keep one agent on the whole task until the work breaks into specialties that need separate context and instructions. Mikiko Chandrasekhar put a practical size on that split: small agents equipped for one to three tasks, each judged on criteria for its own skill set. After the split, log every handoff so you can see who was called, what crossed between them, and what came back.
Mikiko Chandrasekhar is Staff Developer Advocate at MongoDB. Chain of Thought episode 33 aired in July 2025. Giovanna Carofiglio of Outshift by Cisco. Episode 31 aired in July 2025.
When should you split one agent into several?
Split when one agent is being asked to carry specialties you would staff as different jobs. Mikiko described the practice forming for complex work as a system of small agents. Each one is equipped for a small slice, one to three tasks, with criteria for how that agent performs, and then you look at the group.
Stay with one agent while the steps share a skill and you can still read one trace from the task, through the tool call, to the answer. She treats agents as software products. One school of thought, in her account, lets agents run because the models are nondeterministic, and catches errors later in the application layer. The school she follows adapts software engineering and MLOps practice and keeps the system as code you can inspect. A second agent, added before that single trace is exhausted, is another component you now have to debug.
At her lightning talk at the AI Engineer World's Fair, she said basically 80% of the room raised a hand for building an agent, and roughly those same people, maybe one or two fewer, raised a hand for having an agent or multiple agents in production.
And then I asked them like, how many of you have gotten your agents to work, like, straight off the bat, a 100% of the time? No one raised their hands. No one at all.
Mikiko Chandrasekhar, Staff Developer Advocate at MongoDB, on Chain of Thought ep 33
A structured workflow, she said, can still differ on one out of 100 runs, or one out of 20, depending on how complex the task is. Another agent adds a place for that spread to show up.
If you do split, compare configurations of the same specialty before you add a new role. She sketched a hypothetical real-estate system with a listing-research agent whose only job is to crawl listings. An MLS was one source she named. In her example you could stand up five or six of those agents, vary the language model, the reasoning model, and the prompt style, keep the listing data in one store, trace why one configuration performed better, and keep that one.
How do you judge each agent on its own criteria?
Write a different check for each skill, then read the team against the job. One shared score tells you the run failed. It does not tell you which specialist failed.
Instead of having each agent is compared to the same criteria, instead, like each agent has a different criteria that is focused on like their kind of specific skill set.
Mikiko Chandrasekhar, Staff Developer Advocate at MongoDB, on Chain of Thought ep 33
For coding or copilot-style agents, where the output is code, she named signals you can count: accuracy, runtime efficiency, and how often someone accepts a suggestion. Lines of code, she said, is a terrible measure. The pushback she described from a technical leadership track is to measure impact, and sometimes that impact does not encode. Vibe checks are fine up to a point. Past that point she wants quantitative metrics, because you can't improve what you can't measure. The idea that an LM or a VLM doing the reasoning means you skip metrics is, in her words, totally false.
Giovanna put a limit on the same dashboard. Select good metrics, and not too many, or you confuse the people reading them. The set has to capture the dynamics of the multi-agent system. Predictability, for her, means testing the space of solutions these agents work in and checking that similar inputs produce consistent output. Evaluation has to explain how the system works, because these systems are much more complex than normal application.
How do you make agent-to-agent handoffs observable?
Log the payload that crosses between agents, along with the final answer. Giovanna wants a view of the agentic graph: how agents communicate, how data passes from one agent to another, and how tools are called. Without that view, a bad handoff and a weak specialist look the same in the final answer.
... this will give both the developer and the enterprise deploying the application in production the way to understand how the agent communicate, how the data are passed between one agent and another, how the tool are called...
Giovanna Carofiglio of Outshift by Cisco, on Chain of Thought ep 31
You can start that log inside one process. A router picks one narrow agent, records the handoff, calls it, and records what came back.
A minimal sketch to adapt, not a drop-in library.
def route_task(task, agents, trace):
name = pick_agent(task, list(agents))
trace_span(trace, "handoff", {"to": name, "task": task})
spec = agents[name]
result = call_agent(spec, task)
passed = spec.check(task, result) # this specialty's own criteria
trace_span(trace, "return", {"from": name, "result": result, "passed": passed})
return result, passed
pick_agent is your rule for which specialty owns the task. call_agent runs that specialty with its own instructions, and spec.check scores the result on that specialty's criteria. trace_span records who was called, what was sent, what came back, and whether it passed. If the researcher returned a thin brief and the coder still shipped code, the log shows which side to fix. If the brief was fine and the next agent never received it, you have a handoff bug. Tightening the coder's prompt will not touch that.
FAQ
When is one agent the wrong shape?
When the task is several specialties and one set of instructions is holding all of them. Mikiko sized each small agent at one to three tasks, with criteria for that agent's performance. If you cannot name a separate check for a second agent, you do not have a second specialty yet.
Which numbers are worth tracking on a specialist?
For a coding agent she pointed to accuracy, runtime efficiency, and how often a suggestion is accepted. She called lines of code a terrible measure. Giovanna's filter is a small set that shows how the multi-agent system moves, including whether similar inputs stay consistent.
What should a handoff log contain?
The agent you called, the task you sent, and the result you got back. That covers who was called and what data passed. Giovanna's agentic graph also includes how tools are called. If you cannot replay the message, you cannot separate a bad specialist from a bad handoff.
Takeaway
- Stay on one agent while one trace still explains the run.
- Split into small agents when the work is separate specialties. Mikiko sized each at one to three tasks.
- Give every specialist its own criteria. A single shared score hides which role failed.
- Log each handoff: who was called, what was sent, what came back.
- For coding agents, track accuracy, runtime efficiency, and accepted suggestions. Treat lines of code as a bad metric.
The longer explainer, with the episode clips, is on Chain of Thought. It draws on this episode.
Subscribe to the Chain of Thought newsletter for new episodes and write-ups like this one.
Drafted with AI assistance from the episode transcripts.
Top comments (0)