Agent versioning

You made 40 changes this week. One quietly changed what your agent does. Which?

DecimalAI versions every change to your agent’s manifest — so you can diff exactly what moved between the version that behaved and the one that doesn’t.

Start free
MANIFEST CHANGE · support-agentbreaking
v1.0 → v2.0
− compare_competitors   ~ system_prompt   + refund_order
↓↓↓
Production regressions

Live traces break.

300 traces will fail
Skill drift

Quality slips silently.

SkillScore 88 ↓ 79
no alert · no signal
Stale training data

Datasets go invalid.

300 drop1,200 repair
4,000 traces now mismatch
With vs without

The same failures — caught, not shipped.

Without a system watching your manifest, these failures reach production before anyone notices. DecimalAI sits in the loop: the impact report gates your PR, contract checks watch live traffic, and the data lifecycle keeps every trace valid — so the same changes that used to break things now surface the moment they happen.

Without DecimalAI
✗
Regressions ship. Traces break — you hear it from users.
✗
Skills drift. Quality slips with no signal.
✗
Data goes stale. You fine-tune on behavior that no longer exists.
With DecimalAI
✓
Caught on the PR. Impact report before you merge — 300 flagged.
✓
Drift alerted. Contract violation fires the moment it happens.
✓
Data stays valid. Traces re-classified against the new version.
On every PR

See what breaks before you merge.

Every pull request that touches your agent gets a manifest diff and an impact report, posted as a comment in under a second. You see exactly which production traces will keep, repair, replay, or drop — before it ships. No eval suite, no agent run, fractions of a cent per check.

decimal-ai bot commented · now
⚠ Decimal Manifest Impact — support-agent

Removing compare_competitors breaks 300 of 5,000 production traces.

2,800 keep1,200 repair700 replay300 drop
analysis cost <$0.001 · 1,892 traces scanned · no agent run
The manifest

The thing observability tools don’t track.

A manifest is your agent’s structural fingerprint — every tool, prompt, model, skill, sub-agent, and output schema the runtime exposed. Tracing and eval tools watch outputs; they don’t watch the manifest. That’s the difference.

Manifest · support-agent · v2.0
⌗ 7 tools⌗ 3 prompts⌗ 2 models⌗ 4 skills⌗ 1 schema
hash · jcs-sha256 · 9f3a…c07e
Versioning new

Every change to your agent, versioned.

The SDK captures the manifest on every run and hashes it, so each structural change becomes a version you can see, diff, and roll back to — reproducible with the open agentversion CLI.

Version timeline
v1.0
7d
v1.1
5d
v2.0
now
− compare_competitors
~ system_prompt >30%
+ refund_order
Contract drift · in production

Catch what slips past CI.

Some regressions don’t come from a PR — a prompt-injected tool call, an SDK upgrade that silently swapped your model. Declare your contract; violations get flagged in live traffic, not just at merge time.

Contract violations · 12 · 24h
tool_not_declared
delete_user is not in your tool_registry
model_out_of_scope
gpt-4o-mini observed · declared gpt-4o only
Training data

Training data that stays valid.

Every trace is classified Keep, Repair, Replay, or Drop against the current manifest. Repair runs as deterministic transforms — zero LLM cost. One click to export a clean, version-matched dataset.

Dataset build · support-agent
2,800 keep1,200 repair700 replay300 drop
4,000 reusable · zero-LLM-cost repair
One engine

The same diff that blocks regressions ranks every skill.

Versioning is half the story — the engine that watches your manifest also powers the skills registry.

Explore the Skills registry →

Stop catching agent regressions by hand.

Two lines of code. Free tier. Open-source SDK.

# two lines
import decimalai
decimalai.init()
Start free
Open source · MIT licensed