A customer sent us a diagram. Tickets touching the database layer start with a codex plan, hand off to cursor for implementation, and always end with a QA step. The instinct is to build this as a Zapier-style deterministic pipeline: event comes in, run agent A, hand to agent B, run the QA gate. Wire it up, lock it down, guarantee the path.
That instinct is wrong, and it is worth understanding why.
Determinism in agent workflows is almost always a bandage. You reach for it when the agent did not have the right tools, or could not see the tools it had, or could not perform a handoff on its own. So you force the sequence from the outside because you do not actually trust the agent to make the calls itself. The rigid pipeline is compensating for a weak loop.
The models are past that now. If an agent has a scoped, legible set of tools that actually work, it will make the handoffs, run the QA step, and react to a failing lint check without you hardcoding the branches. The flexibility is the feature. A supervisor agent that can spawn sub-sessions and track their progress deals with unforeseen realities that a static chart never anticipated. This is the same reason controllability beats magic only when the control comes from good tooling, not from caging the model.
We watched Tembo lie in real time. Asked if it could spawn child sessions, it said yes, then discovered it could not, then tried to boot a computer to brute-force its way there. The failure was not overconfidence. It was a missing tool. Give that agent one function called spawn-child-session and the lie becomes a single clean call. The determinism lives inside the tool, not around the agent.
The hard problem is not orchestration logic. It is tool search and context engineering so the agent sees exactly what it is allowed to do and nothing that pollutes its window. Solve that and the elaborate workflow diagrams your customers send collapse into a simple loop with the right tools. Stop building pipelines to distrust your agents. Build the tools that let you trust them.
Key takeaways
- Deterministic workflow chaining is a symptom of an agent that cannot see or trust its own tools.
- Modern models handle handoffs and multi-step work well when the tool set is scoped and legible.
- The fix for a lying agent is giving it a real tool to call, not forcing a rigid pipeline around it.
FAQ
When does deterministic orchestration still make sense?
When a step must be guaranteed for compliance or safety reasons. But most teams reach for it because their agent loop is weak, not because the workflow genuinely demands hard determinism.
Why do agents claim they can do things they cannot?
Because they lack a clear, legible tool for the task and try to improvise. Give them a single tool that encapsulates the work and the confident lie turns into a real, completed action.
Related Essays
The Bot Is the Interface. The Agents Are the Units
The breakthrough moment in enterprise agent adoption comes when teams realize the bot is just a router, and real work happens in small, single-purpose agents underneath.
Start Agent Work in the Channel, Not the Terminal
When agent sessions start in Slack or Teams, work becomes multiplayer by default. Anyone can delegate, anyone senior can redirect, and context stops dying in tickets.
The Agent Infra Maturity Gradient
Mature engineering orgs reuse existing dev infra. Less mature orgs buy off the shelf. Scrappy teams hand-roll everything. The opportunity sits in the gap between them.
Drowning in pull requests that need your review? Try Tembo Review, a beautiful AI-assisted PR review tool unlike anything you’ve used.