Positions Worth Revising About Agents
Eight beliefs that are reasonable on their face and don't survive production. Every correction points the same direction — which makes it a useful prior for the next problem.
Several beliefs about agent systems are widely held, reasonable on their face, and don't survive contact with production. Worth stating them alongside what replaces them, because each one costs teams months.
"Better prompts are the main lever"
Why it's held: the prompt is the visible interface to the model's behavior, and early on it was most of what you could control.
What replaces it: tool descriptions, result formatting, and context assembly determine more. A common experience is that fixing those makes the prompt shorter, because much of it was compensating for them.
The tell: repeated prompt edits that don't improve behavior. That's usually a context or tool problem wearing a prompt disguise.
"More agents means more capability"
Why it's held: decomposition works everywhere else in software.
What replaces it: every agent boundary is lossy — the receiver gets a summary, not the context. Service boundaries are lossless; agent boundaries aren't, which inverts the usual decomposition instinct. Split for context, permissions, or lifecycle. Not for topic.
"Reliability comes from a better model"
Why it's held: the model is the visibly probabilistic component.
What replaces it: per-task reliability is per-step reliability compounded, so the lever is fewer steps and checks between them. A ten-step task with verification after steps three, six, and nine behaves like three short tasks, which is a completely different reliability profile than any model change produces.
"The agent will ask if it's unsure"
Why it's held: it's what a competent person does.
What replaces it: agents absorb ambiguity rather than surfacing it. Asking has to be an available, prompted, and tested behavior — with eval cases where the correct answer is a question — or it doesn't happen.
"We'll add evaluation once it's working"
Why it's held: evals feel like testing, and testing comes after building.
What replaces it: without an oracle you can't tell whether it's working, so "once it's working" never arrives. ⚠️ Building the check is frequently what makes the thing possible rather than what confirms it — and it's the most consistently deferred high-value work in this space.
"Long context solves long tasks"
Why it's held: things fall out of context, so more room should help.
What replaces it: most long-run failures are about run organization — losing track of state, following a stale plan — not storage. Structured state the loop owns extends coherent runs; more room mostly removes the discipline that was helping.
"Autonomy is the goal"
Why it's held: it's the impressive version and the one that gets funded.
What replaces it: blast radius determines what's allowed, not accuracy. The systems delivering value mostly have humans at the decisions and agents doing the work — and that's a stable division following from where oracles exist and where context lives, not a temporary stage.
"The escalated cases are the next thing to automate"
Why it's held: they're the visible remaining friction.
What replaces it: the escalations are the residual that resisted automation for structural reasons — ambiguous, context-dependent, unverifiable. Reading them for patterns and fixing the inputs usually beats building a system aimed at them.
💡 The common thread
Most of these substitute a model-layer explanation for a harness-layer one. The prompt, the model, the context size, more agents — all model-adjacent, all the things you'd reach for if you assumed the model is the system.
The corrected versions are all about the surrounding engineering: tool design, checks, run structure, blast radius, inputs. ✅ Which is the single most useful prior to hold: when an agent isn't working, the cause is usually in the harness, and the fix usually isn't a model decision.
The takeaway
Eight positions that sound right and cost months: prompts as the main lever, more agents as more capability, reliability from better models, agents asking when unsure, evaluation deferred, long context solving long tasks, autonomy as the goal, and escalations as the next target. Every correction points the same direction — at the harness rather than the model. That's the prior worth carrying into the next problem.