Agent engineering
HiveMind
A shared layer of understanding for how we build with agents.
176K
coding sessions captured
across 162 engineers · 260 devices · 820 repos
15M+
lines of code written by agents
28,923
sessions linked to a PR
4.3M
tool calls observed
SESSIONS PER MONTH
Explosive growth 💥
73,604
apr 2026
Jan '26Feb '26Mar '26Apr '26
SESSIONS BY AGENT
What it means to develop and maintain software has
fundamentally changed.
01
You can't improve what you don't measure.
No data on agent usage means we're guessing.
02
Sharing skills and techniques should be easy.
Your best workflow shouldn't die in your terminal.
03
Visibility matters more as you run more agents.
More concurrent sessions, more surface area.
03
Who it's for
IC
Engineers who want to get better at prompting.
See how teammates solve similar problems.
STAFF
Staff+ who want to share patterns across teams.
Individual tricks → team practices.
LEAD
Leads and managers who need to
understand their team.
understand their team.
Where's the work? What's blocking?
Two products. Two audiences.
WEAVE
For teams putting AI into their products.
Instrument the AI features you ship: evals, traces, prompts.
user → your AI feature
HIVEMIND
For teams using AI to build their products.
Instrument how your agents write code: sessions, tools, what ships.
engineer → AI agent → code
PART ONE
Product
overview.
overview.
architecture
·
privacy
·
features
Architecture
ON YOUR MACHINE
hivemind daemon
claude code
cursor
codex
gemini
opencode
Discovers sessions. Ships raw JSONL. Redacts secrets.
INGEST
FASTAPI
normalize & store
→ AG-UI events
agentstream adapters
raw + normalized
Raw sessions in. Normalized AG-UI events out.
QUERY
DATA + UI
ClickHouse + React
sessions · turns · tools
PRs · $$$ · trajectories
search index
dashboard
Fast analytics on everything the team has done.
// daemon is intentionally dumb — only discovery, redaction, and syncing
infra: clickhouse · fastapi · terraform
Privacy · 1 of 3
SESSION
GITHUB MATCH
VISIBILITY
01
refactor inference loop
org repo · team has read
TEAM-VISIBLE
Anyone with repo read.
02
deploy hotfix to staging
private repo · 4 collaborators
REPO-LOCKED
Only repo collaborators.
03
draft a board update
unmatched
SOLO
Only you. Always.
// override at any layer
Security · 2 of 3
GATE 01 · DEVICE
Keychain-stored, auto-rotated.
Short-lived tokens. Never on disk in plaintext.
macOS Keychain · libsecret
24-hour rotation
instant dashboard revoke
GATE 02 · IDENTITY
SSO & SCIM, straight from your IdP.
Membership and offboarding flow from Okta. No parallel directory.
Okta · Azure AD · Google
SCIM auto-deprovision
group-scoped access
GATE 03 · NETWORK
Private VPC, fully terraformed.
Data plane behind private connect. No public ingress.
VPC + private connect
infra in version control
audit log on every read
PERIMETER
Pull a token, an SSO group, or VPC peering — the next request stops at the door.
Sensitive data · 3 of 3
// on your laptop
RAW
# trajectory.jsonl — pre-egress cmd : curl -H "Bearer sk_live_8f3d92ab" env : ANTHROPIC_KEY=sk-ant-9a3… DATABASE_URL=postgres://p@ss AWS_SECRET=wJalrXUtnFEMI… diff: + STRIPE_KEY=sk_live_K3y…
// over the wire →
EGRESSED
# trajectory.jsonl — post-egress cmd : curl -H "Bearer [REDACTED]" env : ANTHROPIC_KEY=[REDACTED] DATABASE_URL=[REDACTED] AWS_SECRET=[REDACTED] diff: + STRIPE_KEY=[REDACTED]
SCRUBBED ON-DEVICE
API keys, env vars, auth headers, and high-entropy strings.
COMING NEXT
Org-wide archive policy and enhanced AI PII redaction.
Features
OVERVIEW
Activity feed
Your week — shipped, open, teammates' work.
LIVE
In-flight sessions
Who's mid-session right now. Tail if shared.
INSIGHTS
Trajectory analysis
Where agents fail. What to fix.
LEADERBOARD
Cross-team stats
Who's shipping. Where. (V2 en route.)
USAGE
Cost & tokens
Spend per user, repo, model. Cost per merged PR.
SESSIONS
Full replay & fork
Every turn. Searchable. Shareable. Resumable.
PART TWO
Demo.
live from my laptop · no slides · YOLO
PART THREE
What we've
learned.
learned.
four months of dogfooding · seven lessons
01
Rich trajectories are a data flywheel.
A dataset that didn't exist before.
02
Seeing teammates' sessions is legitimately useful.
"How would Tim do this?" — one click away.
03
Live + Overview unlock parallel work streams.
Tail one session while another runs.
04
The leaderboard is loved — and tricky.
Rewards burn, not outcomes.
05
Fork is a lifesaver when Anthropic is down.
One click to hand off a stuck session.
06
Our bottleneck is QA and code review.
HiveMind can help.
07
The Agent is a User.
hivemind --help FTW.
The thing I'm sure of
RAW SESSION
↓ ENRICHED ON INGEST
PROMPT
user turn
bug
feature
fix
refactor
→
THOUGHT
reasoning
intent
→
TOOL
edit · bash · grep
skill: frontend-design
subagent: code-search
✕ exit 1
→
DIFF
code change
files · langs
→
PR
shipped (or not)
merged
reverted?
where did anyone use the frontend-design skill last week?
Infinite agents. One shape. Already changing how we debug, teach, and build tooling.
Mines a teammate's sessions; produces a sub-agent that talks like them. Built by a user, not us.
$ /soul-stealer tssweeney "Tim Sweeney"
→ spawning hivemind sub-agent…
→ analyzing 47 sessions · 1,284 prompts
→ clustering behavioral modes…
→ extracting negative space (rejections, reversals)…
→ fingerprinting vocabulary…
✓ ~/Desktop/skills/talk-to-tim/SKILL.md
$ /talk-to-tim "should this be a class or a function?"
Tim: Why is this a class? Flatten it. We'll add structure when there's a second caller.
TS
Tim Sweeney
@tssweeney · popular soul
THINKING PATTERNS
REJECTS
FINGERPRINTS
// we did not build this. we built the substrate. the org built the skill.
New feature
overview.hivemind / your-week
// not a dashboard. a memory aid.
The hard one
V1 · TODAY
token count
Rewards burn. Ignores whether the code shipped.
V2 · NEXT
outcomes
Merged PRs · 30-day survival · cost per merged line.
// incentives shape behavior. if we get the scoreboard wrong, the players optimize for the wrong game.
When the model goes down
// we don't control Anthropic's uptime. we do control how fast you keep going.
The new bottleneck
SHIPPED · THUMBNAILS
Skim a session at a glance.
NEW · AI WALKTHROUGH
Read the agent before the diff.
00:42
SUMMARY
Refactored inference loop · 4 files.
02:18
WORTH A LOOK
Silent fallback at batch.py:142.
06:55
SELF-CORRECTED
Reverted breaking change after test failed.
11:03
SKIPPED
Legacy-path tests not run.
// review the trajectory · not the patch
~15 min → ~3 min
// QA is the real bottleneck. we're tooling it.
Built for them, used by them
We design for humans and for the agent reading their terminal. Both want the same thing: a clean CLI.
$ hivemind --help Usage: hivemind [OPTIONS] COMMAND [ARGS]… Hivemind — sync agentic coding sessions to the cloud. Sessions: import Import sessions from local agent history. transcript Fetch and display a session transcript. search Search sessions or list recent ones. fork Fork work from a previous session. insights Manage contextual intelligence suggestions. export Export your data as Parquet or DuckDB.
AUTO-INSTALLED
@hivemind subagent.
Every login drops a Claude Code subagent into ~/.claude/agents. Mention it and it knows your sessions, your forks, your skills.
› @hivemind find sessions where I used the
frontend-design skill last week
→ calling hivemind search …
→ 4 matches · ranked by recency
✓ ready to fork or summarize
// good CLI · readable by both humans and the agent reading their terminal.
PART FOUR
Where
it's going.
it's going.
three bets · next quarter
Where we're going
01
Service agents
Agents that work while you sleep.
Async agents on PR comments, alerts, on-call pages. Federated identity. Sandboxes included.
02
Insights V2
A classifier that earns its keep.
Rebuilt on Weave evals. Repo-wide patterns. Insights → PR in one command.
03
Code quality
Did the agent's code matter?
Track what merges, what survives, what gets reverted. Receipts for the spend.
// each one gets its own slide. let's go.
01
Same shape. Same insights. No human in the loop.
// trigger
An event happens.
- PR comment
- Sentry / pagerduty
- Slack mention
- Cron / schedule
→
// run
An agent runs — no secrets.
- WIF: GH Actions, k8s, Modal
- Short-lived tokens
- Your sandbox or ours
→
// tracked
It shows up in the feed.
- Same trajectory shape
- Same insights, same leaderboard
- Replays, redaction, privacy
// every event becomes a session. every session becomes a learning.
02
Trajectories → "here's what's broken." Today, noisy. Next, measured.
V1 · today
Noisy.
- One-shot, no eval rigor.
- False positives drown signal.
- Single-session issues, not repeating ones.
- Insights you read — not act on.
V2 · next
Evaluated.
- Built on Weave evals.
- Patterns across users in a repo.
- Insights → PR in one command.
- Falsifiable when the fix fails.
// we're eating Weave's dog food. it's the only way this gets better.
03
Generation is cheap. Survival is the metric.
t = 0
generated
agent writes a diff
t + min
PR opened
does it pass CI?
t + day
merged · or not
how much did review rewrite?
t + 30d
still here
replaced? reverted? rewritten?
t + 90d
survived
in prod, shipping value
payoff · engineers
A leaderboard that rewards survival, not burn.
Cost per merged line. Code still in main after 30 days.
payoff · stakeholders
A clear answer to "what is the AI capex buying?"
PRs shipped, code surviving — by repo, team, quarter.
// generated is cheap. survived is the metric.
Great tools are part of our DNA.
HiveMind is how we stay ahead in
a landscape that's accelerating.
HiveMind is how we stay ahead in
a landscape that's accelerating.
Over to you
Questions.
SLACK
#hivemind
REPO
wandb/agentstream
INSTALL
brew install wandb/taps/hivemind
SITE
https://hivemind.wandb.tools