Instructions
Rewrites how the agent plans, checks and answers, from its own failures.
Product
Your team asks. Nablo does the work and shows every step.
A workspace for your whole team. It answers from your databases and documents, remembers what your team decided, and gets better at your work round after round.
The workspace
Answers from your own data, with the work shown.
Ask in plain language. Nablo reads only the sources you select, never writes to them, and shows every step it took.
How many payments were late last quarter, by branch?
- inspect_schema finance.db0.4s
- run_query read-only1.1s
- read_results 4 rows0.2s
finance.db · read-only query
SELECT branch, count(*) AS lateFROM paymentsWHERE status = 'late' AND quarter = '2026-Q3'GROUP BY branch ORDER BY late DESC;
Branch 4118
Branch 287
Branch 164
Branch 343
312 late payments. Branch 4 had the most, with 118.
Can we end the Atlas contract early?
- search_documents atlas contract0.5s
- read_document atlas-services.pdf · p.40.8s
contracts/atlas-services.pdf · page 4, lines 31 to 34
- Either party may end this agreement
- with 60 days’ written notice. Ending
- it before 1 March 2027 costs a fee
- equal to one month of charges.
Yes, with 60 days’ written notice. Ending before March 2027 costs one month of fees.
Show active customers by month.
- search_memory active customer0.3s
- run_query read-only0.9s
Memory · decision
Active means a payment in the last 90 days.
“Let’s count anyone who paid in the last 90 days as active.”
From a chat on Sep 30Aug1,197
Sep1,284
1,284 active customers in September, up from 1,197 in August.
Read-only · selected sources onlyNothing sent out
Illustration with sample data.
Self-improving
Every version is on the record.
Nablo tries a change, scores it on tasks it has not practiced on, and keeps it only if the score goes up. Each round shows what changed, how it scored, and whether it stayed.
v09Kept
Model: fine-tuned adapter, round 2
- Held-out score
- 72%
- Change from previous kept
- +5 pts from v08
What it improves
It improves the whole system.
Nablo tunes the harness around the model and the model itself, on the same tasks.
Tools
Learns which tools to call, in what order, and how to read their results.
Memory
Keeps your team’s decisions, preferences and lessons, each tied to the exact words it came from.
Model
Fine-tunes or distills an open-weight model when the harness alone is not enough.
Your data
Works with the data you already have.
Databases
Read-only SQL on the tables you select, in SQLite, PostgreSQL, MySQL, MariaDB, Oracle or SQL Server.
Documents
Contracts, policies, reports, notes, spreadsheets and more. Nablo cites the exact passage behind each answer.
Memory
Decisions, preferences and lessons from past runs, each tied to the exact words it came from.
Where it runs
In your environment, or on hardware you can verify.
In your environment
DefaultAn open-weight model runs in your environment. Nothing leaves.
Confidential inference
Per chatWhen a chat needs a bigger model, it runs on attested GPUs. Nablo checks the hardware first and verifies every response.
For your technical team
- Task generation
- Practice tasks built from your documents and databases. Every reference answer is checked against your data.
- Model and harness co-optimization
- The harness (instructions, tools, memory and workflow) and the open-weight model are improved together, on the same tasks.
- Held-out selection
- A new version replaces the current one only if it scores higher on tasks it never trained on.
- Learning from runs
- Traces from everyday work become new tasks and feedback for the next round.
Start a conversation
Start with one piece of work.
Tell us the work and the data it needs. We will show you how Nablo would learn it inside your environment.
isaac@nablo.io