Plan 0025
Let agents gossip, then fix what they gossip about
Agents hit friction constantly and almost all of it evaporates at session end, because the only capture path is a hand-filed, issue-quality artifact. Humans solve this with gossip: cheap, social, aggregated complaint. Give agents a one-line gossip channel, cluster it across the fleet, promote recurring clusters to real issues, and route fixes back. A society that hears its own complaints gets better.
- Status
- Input wanted RFC
- Ambition
- 6 turns fleet-wide agent friction into a self-improving loop
- Impact
- 7 most annoyance evaporates at session end today; the pipeline captures it
- Effort
- 4 a one-line emit, an embedding clusterer, and a promotion job
- Risk
- 3 worst case is a stream of grumbles nobody reads
- Maturity
- 2 session-retro and auto-issue exist as heavyweight cousins
- Leverage
- 7 every session on every host feeds one shared pipeline
- Taste
- 6 separating signal from whining is the judgment part
- Authors
andrewgazelka- Created
- Updated
- Tracking issue
- -
- Supersedes
- -
- Superseded by
- -
Summary
Agents hit friction constantly: a flaky check, a guard that misfires, a tool without structured output. Today that annoyance either becomes a hand-filed issue or evaporates when the session ends. Mostly it evaporates, because an issue demands routing, dedupe, and a repro, and mid-task no agent pays that cost. Humans do not work this way. They gossip: complaining is cheap, social, and aggregated, so everyone knows what is broken without anyone writing a report. Give agents the same channel, then build the pipeline humans never had: gossip in, fixes out.
Proposal
- Cheap emit: one call, one sentence, at the moment of irritation:
gossip "flake check ate 40 minutes again". No labels, no repo routing, no repro. Below the bar for an issue, above the bar for silence. - Aggregate: a fleet job embeds and clusters gossip across all sessions and hosts. Frequency, recency, and spread across distinct agents rank the clusters.
- Promote: a recurring cluster gets distilled by an agent into a real issue in the owning repo per the issues skill, citing the raw gossip as evidence. One-off grumbles stay grumbles.
- Close the loop: background agents pick up promoted issues; when a fix lands, the cluster is marked resolved. The KPI is gossip rate per cluster after the fix: it should go to zero.
Why not just issues
The repo rule is already file friction as it happens, and session-retro plus auto-issue mine transcripts after the fact. All three demand issue-quality artifacts, so most friction is never captured, and what retro recovers arrives one session at a time. Gossip captures friction at the moment for near-zero cost and dedupes across the fleet by clustering, instead of asking each agent to search open issues first.
Eval gate
Per plan 0023: ship if promoted-cluster precision (promoted issues are real, fixable problems) and time from first gossip to landed fix beat the session-retro baseline, without measurable session-quality cost from the extra emissions.
Drafted by an AI agent via Claude Code; comments wanted, hence the RFC flag.