Letting the Agent Propose Its Own Intents, and Reviewing Them Before They Ship
An agent that notices a gap in what it can already do, proposes a fix, and waits for a person to approve or reject it — a narrow, concrete instance of 'agent proposes, human decides.'
"The agent proposes, the human decides" is the general shape that most 2026 human-in-the-loop thinking converges on as the right balance of autonomy and oversight. It's easy to nod along with as a principle. It's more interesting as a specific, real feature — and Piskes has one that applies that exact pattern to a genuinely useful, narrow problem: an agent noticing its own gaps.
What an intent suggestion actually is
As your agent handles real conversations over time, it can notice a recurring pattern it doesn't currently have a clean, named way to handle — the kind of request that keeps showing up in a similar shape but isn't yet formalized as something the agent explicitly knows how to do. Rather than silently improvising a slightly different answer every time, it can propose a new intent: a suggestion, sitting in its own review queue, for a capability the agent thinks it should formally have.
That suggestion doesn't take effect on its own. It sits, waiting, until a person looks at it and either approves it — turning the pattern into something the agent formally handles going forward — or rejects it, deciding the pattern either isn't real or isn't something worth formalizing.
Why this is a different thing than general guardrails
It's tempting to file this under "guardrails" and move on, but it's actually a distinct pattern worth separating out. Guardrails are about constraining what an agent is allowed to do. Intent suggestions are about the agent contributing to what it's capable of doing in the first place — but only ever as a proposal, never as a fait accompli. The agent gets a voice in noticing its own gaps; it doesn't get to act on that observation unilaterally.
That's a meaningfully narrower and more concrete instance of "agent proposes, human decides" than the general principle usually gets articulated as. It's not a philosophy statement. It's a specific queue, a specific approve/reject mechanism, and a specific trigger — a pattern the agent noticed in conversations that actually happened.
Why it's worth using
If you're running an agent on any kind of volume, it's going to encounter requests it doesn't have a clean, named way to handle more often than you'd expect — and by default, it'll just improvise something reasonable-but-inconsistent each time rather than flag the gap. Intent suggestions turn that invisible drift into something visible and reviewable: instead of a slowly accumulating pile of ad-hoc improvisation, you get a queue of specific, nameable proposals you can approve, reject, or refine — and the agent gets progressively better at handling your actual, real traffic instead of staying frozen at whatever it shipped with.