You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
@@ -17,7 +17,26 @@ The "AI teammate" framing creates a different mental model entirely:
17
17
18
18
### Inspiration: OpenClaw & the "Trainable Agent" Pattern
19
19
20
-
OpenClaw and similar projects (SWE-agent, Devon, etc.) demonstrate that the most powerful AI agents aren't the ones with the most tools — they're the ones that **learn from their environment**. The key insight:
20
+
**OpenClaw** (247K+ GitHub stars, fastest-growing open-source project ever) proved the "teammate" framing works when backed by real architecture. Key lessons:
21
+
22
+
1.**Meet users where they are.** OpenClaw's UX *is* your existing messaging apps (WhatsApp, Telegram, Slack, Signal). Zero learning curve. For altimate, the equivalent: meet data engineers in their terminal, their dbt workflow, their Slack — don't force them into a separate app.
23
+
24
+
2.**Self-improving memory.** OpenClaw captures learnings, errors, and corrections in structured files (`LEARNINGS.md`, `ERRORS.md`). When patterns recur 3+ times across 2+ tasks within 30 days, they auto-promote into permanent system prompt files (`CLAUDE.md`, `SOUL.md`). This is the model for altimate's training system — learning should be automatic, not manual.
25
+
26
+
3.**Proactive heartbeat.** A scheduler wakes the agent at intervals so it can act without being prompted — checking email, running tasks, flagging issues. For altimate: imagine the teammate running nightly cost checks, freshness monitors, or schema drift detection without being asked.
27
+
28
+
4.**Persistent identity.** One agent instance across all channels with shared memory and context. For altimate: the same teammate across TUI, web, CI/CD, and Slack — always knowing your project, your standards, your history.
29
+
30
+
**Devin** ($10.2B valuation, $73M ARR) proved the market appetite: they market as "a collaborative AI teammate" and "the first AI software engineer," but candidly advise treating it as "a junior coding partner." The honesty works — users understand the capability boundary.
31
+
32
+
**Factory AI** positions autonomous "Droids" that embed into existing workflows (VS Code, JetBrains, Slack, Linear). Their insight: "delegate complete tasks like refactors, incident response, and migrations without changing your tools."
33
+
34
+
The **World Economic Forum** outlines the industry evolution:
35
+
1.**Copilots** (assisted intelligence) — suggestions, human controls
altimate should skip straight to level 3 for data engineering.
21
40
22
41
> The best AI teammate is one that adapts to YOUR team, not one that forces your team to adapt to IT.
23
42
@@ -100,7 +119,7 @@ patterns:
100
119
- naming: snake_case
101
120
```
102
121
103
-
#### b) Learn-by-Correction (`/feedback`)
122
+
#### b) Learn-by-Correction (Implicit + `/feedback`)
104
123
105
124
When the teammate does something wrong, instead of just fixing it, you teach it:
106
125
@@ -125,6 +144,18 @@ applies_to:
125
144
severity: high
126
145
```
127
146
147
+
**Auto-promotion (inspired by OpenClaw's self-improving agent):**
148
+
149
+
When a correction pattern recurs 3+ times across 2+ sessions within 30 days, it auto-promotes from a "soft preference" to a "hard rule" in the system prompt. The teammate proactively tells you:
150
+
151
+
```
152
+
Teammate: I've noticed you've corrected my use of LEFT JOIN → INNER JOIN
153
+
in staging models 4 times now. Should I make this a permanent rule?
154
+
[Yes, always use INNER JOIN in staging] [No, keep case-by-case]
155
+
```
156
+
157
+
This mirrors OpenClaw's `LEARNINGS.md` → `CLAUDE.md` promotion pipeline, but adapted for structured data engineering rules rather than free-form text.
158
+
128
159
#### c) Learn-by-Document (`/train`)
129
160
130
161
For more formal training, point the teammate at documents:
@@ -315,7 +346,54 @@ Teammate: Learned! This pattern is saved to .altimate/learned/patterns/staging-m
315
346
Commit this file so your teammates' altimate instances learn it too.
316
347
```
317
348
318
-
### 6. The Training Dashboard (`/training-status`)
349
+
### 6. Proactive Behaviors (inspired by OpenClaw's Heartbeat)
350
+
351
+
A real teammate doesn't wait to be asked. OpenClaw proved this with its heartbeat scheduler — a background process that wakes the agent at intervals to act autonomously.
352
+
353
+
For altimate, proactive behaviors are domain-specific and high-value:
354
+
355
+
#### Background Monitors (opt-in)
356
+
357
+
```yaml
358
+
# .altimate/monitors.yml
359
+
cost_alert:
360
+
schedule: daily
361
+
threshold: 120% # Alert if costs >120% of 30-day average
The unique combination: **trainable + data-domain-specific + warehouse-connected**.
534
+
### What altimate takes from each:
456
535
457
-
No other product lets you teach an AI your team's SQL standards and then have it enforce those standards with direct access to your warehouse metadata, lineage, and cost data.
536
+
| From | What we borrow | How we adapt it |
537
+
|---|---|---|
538
+
|**OpenClaw**| Self-improving memory, auto-promotion of learnings | Structured YAML rules instead of free-form markdown; domain-specific (SQL patterns, not general tasks) |
0 commit comments