Context infrastructure
Context infrastructure
for reliable AI agents.
Before an AI agent acts, GammaLex gathers the information the task depends on from your systems, checks it for gaps and conflicts, and tells the agent whether it has enough to proceed — with the evidence attached.
Research · Industry · Public innovation
- NMAICNew Mexico Artificial
Intelligence ConsortiumNMAICNew Mexico Artificial
Intelligence Consortium
Know if your agent has enough context to act.
One call to GammaLex. Context, evidence, and READY, WARN, or BLOCK back.
The agent asks GammaLex
One call before it reasons or acts. It sends the task, who it is acting as, and what the task concerns.
Context request
context = glx.get_context(
- task"Evaluate whether transaction #3812 has suffi…"
- actor"finance_agent_prod"
- scope{ id: "transaction:3812" }
)
One call, three arguments — the agent has not opened a record yet.
GammaLex determines what the task requires
It works out what information should be present and what the agent is permitted to use before relying on any of it.
Permission boundary
- 12msActor resolvedfinance_agent_prod
- 18msScope resolvedtransaction:3812
- 31msPermission boundary applied2,481
2,481 records permitted · 18,204 excluded at the boundary
It assembles the context
It gathers the relevant records and follows how they relate, preserving where each came from and why it belongs.
Selected context
- 01Transaction #3812PostgreSQL
- 02Invoice #8831QuickBooks
- 03Applicable PolicyPolicy Repository
- 04Regulatory RequirementSharePoint
- 05Prior DecisionGoogle Drive
17 records selected · 24 relationships kept · a reason for each
It checks whether the context is sufficient
Missing, conflicting, changed, unverified, or out-of-bounds information is caught before the agent relies on it.
Findings
- × MISSINGRequired supporting evidence was not found.
- × CONFLICTINGTwo sources disagree on classification.
- ! CHANGEDApplicable policy changed after the underlying record was created.
Found before the agent relied on any of it.
It returns context and a verdict
The agent gets the evidence with READY, WARN, or BLOCK, then decides whether to reason or act. GammaLex does not act for it.
Context package
- context17 records, with provenance
- findings3, with the evidence behind each
- readiness× BLOCKED
Context is not ready for autonomous execution. Required supporting evidence is missing.
One complete run.
Which records GammaLex selected, what it found wrong, and why it stopped the agent — exactly as the agent receives it.
GammaLex AIContext / Runs / RUN_8F92ABContext Run
RUN_8F92AB
Evaluate whether transaction #3812 has sufficient context for the requested decision.
Sep 8, 2026 · 10:24 AM
- Requested by
- finance_agent_prod
- Via
- SDK
- Environment
- Production
- Scope
- transaction:3812
Context is not ready for autonomous execution. Required supporting evidence is missing.
17 records selected across 5 sources·3 findings·2 blocking
5 of 8 sources connected
Every check ends with one of three answers.
READY
The agent has everything the task needs. It is current, consistent, and permitted.
The agent proceeds.
WARN
Usable, but something is out of date or unsettled.
The agent proceeds. The concern travels with the output.
BLOCK
Something the task needs is missing, contradicts another record, or is not permitted for this agent.
The agent stops. The reason is returned instead of a guess.
Five reasons the answer is not READY
- Missing
- Something the task needs was never there.
- Conflicting
- Two records say different things.
- Changed
- Something it depends on changed after the fact.
- Uncertain
- What is there does not settle the question.
- Boundary
- Something relevant exists, but the agent is not permitted to use it.
What GammaLex gives an agent.
The relevant information
Only the records this task depends on, drawn from what the agent is permitted to read — with the reason each was included.
A clear readiness decision
READY, WARN, or BLOCK before the agent acts — and the specific reason: missing, conflicting, changed, uncertain, or out of bounds.
Enforced permissions
Information the agent is not permitted to see never reaches it. Relevance never overrides access, and every exclusion is recorded.
A record of every run
What was used, what was left out, what had changed, and why. Any run can be replayed later against current data.
Not search. Not guardrails. The step in between.
RAG / vector search
- It does
- Finds what looks relevant.
- It can't
- Say what is missing.
GammaLex
Follows the records the task depends on and reports the one that should exist but does not.
Why it is its own layer
- The check can’t come from the same model that is about to act.
- Access rules are defined by your systems, so GammaLex enforces them at the source — not in a prompt.
- Corrections and prior runs accumulate over time, independent of the model or framework in use.
Built for the systems your agents already use.
Your agent makes one call. GammaLex reads from the systems the task depends on; it never writes to them.
- SDK
- Python and TypeScript. One call returns the information and the readiness answer.
- API
- POST /v1/context from any runtime. Same request, same response.
- MCP
- Four tools an agent can call directly, inside the framework it already runs in.
Source adapters
- PostgreSQL
- Snowflake
- QuickBooks
- Google Drive
- SharePoint
- S3
- REST API
- Policy Repository
PostgreSQL
Snowflake
QuickBooks
Google DriveIllustrative adapters — availability varies by integration.
MCP tools
get_contextAssemble and return task-ready context.inspect_contextInspect a previously returned context package.get_evidenceRetrieve the evidence behind a finding.replay_contextRe-run a prior Context Run against current sources.
Open source · OpenContextually · In development
The same check, for coding agents. OpenContextually decides what an agent should read before it touches a codebase. No LLM, no network, deterministic — and it never edits your code.
One context layer.Different environments.
The sources change. The need for reliable context does not.
Reference implementations
Designed for context-intensive environments across financial systems, governance, enterprise workflows, and beyond.
Building agents for real-world workflows?
Work with GammaLex to test the context layer against your systems, requirements, and agent workflows.
Give your agents context they can rely on.
Taking on design partners running agents in high-consequence workflows, and teams integrating the layer into their own product.