Build
Escalation that does not lose context
Tobiloba Odejinmi · 2 Jun 2026 · 6 min · 1,044 words
Direct answer
Escalation that does not lose context is a handoff of working state plus a decision brief. Not a raw transcript. Agent-to-agent triggers and agent-to-human triggers are written separately. Hard gates stop the employee from acting. After a person decides, there is a reentry path so the work does not die in a side thread. On a healthy first process, about 10 to 15 percent of cases escalate.
- Handoff equals working state plus a decision brief.
- Agent-to-agent and agent-to-human triggers are different exits.
- Hard gates are rules, not tone guidelines.
- A 10 to 15 percent escalation rate is a sign the slice is real.
What should an escalation include?
Working state and a decision brief. Working state is the record: who, what, which tool IDs, which fields are known, which are missing, what has already been tried. The decision brief is the question the person must answer, the options, and the consequence of each. If you cannot write the question in one sentence, the employee is not ready to escalate. It is still confused, and you should not donate that confusion to a human.
I do not send a chat log and a shrug. I have sat in incident reviews where the useful fact was three screens down. Escalation is the same failure mode at smaller scale. The person who picks up the case should be able to act without becoming an archaeologist.
When should an agent hand off to another agent?
When the next job is still mechanical and the payload is clean. A classifier that routes a billing ticket to a billing workflow. A research step that returns sources to a drafting step. Those are agent-to-agent triggers. They need a schema, a timeout, and a rule for what happens when the second agent fails. They do not need a person in the middle if no hard gate fired.
What they must not do is wander. If agent B can send work back to agent A without a bound, you built a loop. Loops look busy and lose the customer. The map writes a one-way or a counted bounce, then a human.
When should an agent hand off to a human?
When judgment, liability, or relationship risk shows up. Those triggers are written on the map as entry conditions, not as 'if you are unsure.' Models are unsure in the wrong places and confident in the worse ones. Unsure is a score you can use. It is not the only gate.
Agent-to-human triggers I write down in week one: policy exceptions, identity that does not match, amounts over a line, legal or safety language, VIP or referral flags, and any action that spends money or makes a promise. Support, hiring, and follow-up all have a version of this list. If your list is empty, you are not being brave. You are being late.
What are hard gates?
Hard gates are stops. The employee cannot complete the action. It can prepare the brief. It cannot send the refund, publish the reply, reject the candidate, or book the slot. Gates are implemented in code and in the tool permissions, not only in the prompt. A prompt gate is a suggestion. A permission gate is a fact.
I learned the expensive version of this on systems that moved money. If a row is wrong, someone does not get paid. An AI employee that can move money or make a medical or legal claim without a gate is not an employee. It is an unattended intern with production keys. Days 4 to 6 of the week include wiring the gate, not just talking about it.
How does the case come back after a person decides?
There is a reentry path. The person records a decision in the same tool: approve, deny, ask the customer, rewrite, or kill. The employee then finishes the mechanical tail, or it stays off the case. What you must not do is leave the decision in Slack and hope someone pastes it back. That is how context dies the second time.
Reentry needs the same working state the person saw, plus their decision, plus who decided. If the employee starts from a blank ticket, you paid for a human and then threw away the work. The handover doc has this path in it. If it does not, the docs are a tour, not an operating note.
What escalation rate should you expect?
On a first process I expect about 10 to 15 percent. That is enough judgment left in the slice to be worth a person, and enough volume left for the employee to pay for the week. I watch the rate in production. A sudden drop can mean the gates were loosened. A sudden jump can mean the incoming work changed, or the lookup broke, and everything looks 'weird.'
Do not target zero. Zero is how you get silent misses. Do not brag about a high rate either. If half the queue needs a person, you picked a process that is still a meeting. Pick a narrower pile, or keep the work human and spend the $5,000 on a process that is actually ready.
Questions people ask
Why not just pass the full transcript?
Because the person then redoes the work. They have to find the customer, the amount, the rule, and the question. That is not a handoff. That is a second shift.
What is a hard gate?
A condition the employee may not cross even if it is confident. Refunds over a line. Legal threats. Medical advice. Promises about dates you have not confirmed. The action stops and a person is named.
Is 10 to 15 percent escalation a failure?
No. It is the band I expect on a first process that still has judgment in it. Zero usually means you are hiding misses. Fifty usually means you picked the wrong slice.
Can one agent escalate to another agent?
Yes, when the second agent has a narrower job and the payload is structured. That is not a substitute for a human on a hard gate.
What if the person never answers?
Then you do not have an escalation path. You have a hole. The map names an owner and a backup, and the monitor watches age. A silent queue is worse than no employee.
Written by
Tobiloba Odejinmi
Head of Engineering at 10mg Health. I have run engineering at Zeeh Africa and sold Insurpass and Shopl. I still write the code. If you have one process that still runs on people copying things, we can look at it in thirty minutes.
