Research that holds up
Research is where agents most often embarrass themselves — confident prose, invented citations, sources that dissolve on click. Sonderr's research stack is built against exactly that: web access that is bounded and read-only by design, a crew that investigates without write power, and a synthesis pass that would rather report a gap than paper over one.
Sonderr ships with bounded, read-only web access built in — no setup, no keys, and no pretending the open web is a safe place. Two filters do the guard-rail work: a verified-domain filter and a blocked-private-hosts rule that stops the agent from knocking on internal addresses. Within that fence, research is real: fetch pages, read sources, cite what was actually read. The mechanics are in Web research & live sources; this guide is the method on top of them.
1 · The research crew — read-only by role
For anything deeper than a quick lookup, open a collaboration room: one lead with write access, joined by six read-only workers who investigate and report. The split is deliberate — workers who cannot write cannot "fix" evidence to match a conclusion. Each worker takes a slice of the question; findings and risk flags come back to the lead, who publishes the verdict. Personas on the roster are labeled roles, not second models with special powers — the UI says so and the docs insist on it. The full room mechanics are in Agent teams & collaboration rooms.
2 · The synthesis pass — the honest verdict
Every room ends with a structured synthesis pass from the lead: agreements (where sources and workers line up), contradictions (where they fight — kept, not smoothed away), gaps (what could not be established), and up to three recommended next steps. Two properties make it trustworthy. It is an AI assessment, not a permission slip — the label is on the output, and you stay the decider. And it is allowed to be incomplete: a synthesis that reports "the sources disagree and we could not resolve it" did its job; one that hides the disagreement did not.
3 · The research brief — a shape you can reuse
The recipes library includes a full research brief workflow: frame the question, assign the room, collect findings with links, resolve or report contradictions, and deliver a document where every claim carries its source. The structure is deliberately boring — question, method, findings, contradictions, gaps, next steps — because boring structures are auditable. Steal it from Recipes — eight real-world workflows.
| Question type | Reach for | Why |
|---|---|---|
| Quick factual lookup | Ask + web access | One source check, one answer, receipts attached. |
| Comparison or state of the field | Research room | Six read-only workers cover slices in parallel. |
| Before changing something critical | Pre-change investigation | The recipe that maps blast radius before Build. |
| Ongoing topic | Saved room + briefs | Rooms restore intact; the brief structure re-runs. |
4 · Honesty rules, applied to every claim
- No invented citations. A source either was fetched during the task or it is not a source. "I recall reading" is not evidence and is never formatted as if it were.
- Uncertainty gets words, not vibes. Measured, estimated, assumed and unknown are distinct labels; the synthesis uses them distinctly.
- Contradictions survive. Two sources that disagree are reported as a disagreement with both links — not averaged into a confident mush.
- Recency is stated. Where a fact has a shelf life, the report says what was seen and when.
5 · Keep the good parts
Rooms are saved and restore intact, so a standing research beat — competitors, regulations, a moving technical field — becomes a reusable crew with a memory of its own framing. Sessions carry the findings forward; Sessions, checkpoints & long tasks explains what survives and what does not. And when research feeds action, the hand-off is the pre-change investigation recipe: map the terrain, then let Build move — with checkpoints, as always. The full arc from question to change is Planning a project end to end.