Changelog
Approvals that expire, and a queue worth reading
The approval queue was becoming a second inbox. It now defends your attention instead of competing for it.
Teams running Gaia for a full quarter told us the same thing: the agents were right often enough that reviewing every action stopped being a check and started being a chore. So the queue now groups, expires, and gets out of the way.
Placeholder recording. Swap for the grouped-approvals capture. The grouping is not cosmetic. An approval queue that grows faster than you can read it stops being a control and becomes a rubber stamp, which is worse than no review at all. It produces the paperwork of oversight without the fact of it. Collapsing by action class means the decision you make is the decision you meant to make.
Improvements
- Actions of the same kind collapse into one row you can approve or reject together.
- Any action class you have approved twenty times in a row now offers to move itself to the log.
- Pending approvals expire after seven days instead of accumulating silently.
- Rejections carry an optional reason, and the reason feeds back into the pattern the agent learned it from.
Fixes
- Fixed the queue badge counting actions that had already expired.
- Fixed keyboard navigation skipping the last row in a grouped set.
Agents close the loop
Every action an agent takes now writes back to the tool it came from, and shows you the receipt.
Until now an agent could nudge a deal and log the touch, but the write-back was best effort. If the CRM rejected the update you found out later, in a report. Actions are now transactional: they either land in your system of record or they surface as a failure you can retry.
Placeholder recording. Swap for the real action-log capture. Transactional write-back turned out to be most of the work. A CRM can refuse an update for reasons that have nothing to do with the agent: a required field nobody fills in, a validation rule written years ago for a process that no longer runs, a record locked by a workflow that fires on save. Each of those now surfaces as a named failure with the offending field attached, rather than as a success that quietly did nothing.
Improvements
- Every agent action carries a receipt: what it read, what it decided, and where it wrote.
- Failed write-backs now retry with backoff and raise a single grouped notification instead of one per attempt.
- The action log is filterable by agent, deal, and outcome.
- Approvals expire after seven days rather than sitting in the queue indefinitely.
Fixes
- Fixed touches being logged twice when a thread was replied to from two mailboxes.
- Fixed the forecast card showing a stale confidence figure for up to an hour after a sync.
One shared memory
Deals, touches, and outcomes now land in a single graph that every agent reads from.
Agents used to each keep their own view of a deal. That worked until two of them disagreed. There is now one memory underneath: connectors write into it, agents read from it, and what an agent learned on Monday is available to a different agent on Friday.
Placeholder image. Swap for the memory graph screenshot. Shared memory changes what a suggestion can be built from. Before, an agent drafting a follow-up knew about the thread it was replying to. Now it knows the account has gone quiet twice before at this stage, that both recoveries came from reaching a second stakeholder, and which of the two messages actually got answered.
Placeholder recording. Swap for the graph-building capture. Improvements
- Ninety days of history are mapped on connect, so the graph is useful before you touch anything.
- Pattern confidence is shown alongside every suggestion, and you can see the deals it was learned from.
- Warehouse sync moved to incremental reads, so a full refresh no longer blocks the queue.
Misc
- Connector status moved into the top bar so a broken sync is visible without opening settings.
Connect in minutes
CRM, inbox, calendar, and warehouse in one pass.
Setup was four separate flows with four separate failure modes. It is now one, and it tells you what it is going to read before it reads anything.
gaia connect --crm hubspot --inbox gmail --warehouse snowflake # → mapping 90 days of history… # → 1,204 deals · 87 patterns learnedPlaceholder recording. Swap for the connect-flow capture. Scopes are listed per tool before you grant anything, in the plain language of what will be read rather than the provider's own naming. If a connector wants more than the mapping needs, that is visible at the point of decision instead of in an audit six months later.
Improvements
- Sixty-plus GTM tools connect without a custom field mapping step.
- Read scopes are listed before you grant them, per tool.
- A failed connector no longer blocks the rest of the setup.
Fixes
- Fixed calendar events from shared team calendars being attributed to the wrong owner.
- Fixed setup stalling when a warehouse role had read access but no warehouse assigned.
Receipts on every action
You can now see what an agent read, what it decided, and why, for every action it has ever taken.
An action you cannot explain is an action nobody keeps turned on. Every agent action now carries a receipt: the records it read, the pattern it matched, and the one-sentence reason it acted. Receipts are permanent and searchable.
Placeholder image. Swap for the action receipt screenshot. Improvements
- Receipts link straight to the source records, so you can check the agent's reading rather than trusting it.
- Confidence is shown alongside the reason, with the deals the pattern was learned from.
- Receipts survive the deal. Archiving no longer takes the history with it.
The reasoning line is written for a rep, not an engineer. It is the sentence someone reads when they want to know why a deal got nudged on a Tuesday, and it is the difference between an agent people leave switched on and one they quietly disable.
Fixes
- Fixed reasoning text being truncated mid-sentence on narrow screens.
Forecast confidence, and the gap under it
The forecast now reports two numbers: what is verified, and what is inferred.
A single confidence figure hides the thing you actually need in the room: how much of the number somebody confirmed in writing, and how much the model inferred from activity. Those are now separate, everywhere the forecast appears.
Placeholder recording. Swap for the forecast-split capture. The gap between the two numbers is the useful one. A quarter where verified pipeline is thin and inferred pipeline is fat is not a confident quarter, however good the headline percentage looks, and now that is visible on the slide instead of discoverable in week eleven.
Improvements
- Verified and inferred pipeline are reported separately, never summed into one figure.
- Every inferred deal shows what the inference was drawn from.
- Forecast history is snapshotted weekly, so you can show what moved and when.
Misc
- Forecast exports carry both figures and the snapshot date.
Ninety days of history on connect
Gaia now maps your last ninety days before it suggests anything at all.
Starting with an agent means trusting a system that has never seen your motion. Starting with the map means you can check its understanding before it is allowed to participate. Connect now triggers a ninety-day read, and the result is something you review rather than something that runs.
Improvements
- History mapping runs on connect and reports what it found before anything is enabled.
- Patterns it learned are listed with the deals behind them, so a wrong read is visible early.
- Agents stay off until you turn them on. Connecting is not consent to act.
Fixes
- Fixed mapping stalling on mailboxes with more than 50,000 threads.
- Fixed duplicate contacts when the same person appeared under two email addresses.