Platform · Research

Research agents

Every serious design decision starts with ground truth. Agentriq's research agents run a reason-and-act loop across foundry PDKs, standards, and the primary literature, then hand back a grounded brief with every claim traced to its source.

01The problem

Foundries like AIM Photonics, imec, and GlobalFoundries each publish their own process design kit, and no two agree. Before an engineer lays out a silicon-photonics circuit, they have to compile ground truth from all of them: a recent low-loss silicon strip waveguide can run near 1.48 dB/cm, a 3 dB coupler has to hold its splitting ratio across the foundry's over-etch and under-etch window, and every design rule shifts from one process to the next. Get a single number wrong and it is a wasted mask set and a lost fabrication cycle.

The foundry ground-truth brief. An example from silicon photonics: one design decision pulls facts from many disconnected sources into a single grounded brief.

02How we solve it

Reason, act, and verify every claim.

Agentriq's agents plan a research question into sub-queries, act through real tools to gather evidence, and reflect on what they find until the brief holds up.

The reason-act-reflect loop. Each cycle pairs a reasoning step with a real tool call, then reflects before the next action, so unsupported findings never reach the brief.

ReAct (reason + act) loop

Interleaves a reasoning step with a real tool call so each retrieval is chosen on the fly and grounded in returned evidence, which keeps the trajectory factual where pure chain-of-thought drifts.

Reflexion and self-reflection

The agent critiques its own draft against the task, writes the lesson to episodic memory, and retries, the loop that lifted HumanEval pass@1 to 91 percent over the prior 80 in the original work.

Tool use and function calling

Structured, typed function calls let the agent query PDKs, standards, and literature APIs with precise arguments instead of free-text guesses.

Query planning and decomposition

A broad research question is broken into ordered sub-queries so each foundry rule, loss figure, and coupler tolerance is retrieved and checked independently.

03What it produces

One grounded brief, every claim cited.

Grounding beats ungrounded reasoning. Published benchmarks show that pairing reasoning with real tool actions raises task success and cuts hallucination.

Agentic research does not stop at finding papers: it can generate, compare and evaluate scientific hypotheses. Published work reports that a multi-agent AI co-scientist reached in about 48 hours a hypothesis about how bacteria develop antimicrobial resistance that a research team had taken close to ten years to establish. The everyday output here is a source-traced brief that reconciles every waveguide loss, coupler tolerance and design rule against its primary source, built on the reason-and-act grounding that raised ALFWorld task success by 34 points over imitation learning.

04The agentic loop

Inside the closed research loop.

Agentriq wraps ReAct and Reflexion in a closed loop: the agent plans, gathers, self-checks, and writes a claim into the brief only once it survives verification against its source.

From work to a verified decision. The research agent plans, gathers, self-checks, judges, and writes a source-traced brief back to memory.

01Plan the question into ordered sub-queries and act through PDKs, standards, and literature APIs.
02Verify each retrieved figure against its primary source and flag anything unsupported.
03On a conflict or gap, run root-cause analysis and pull targeted evidence with Agentic RAG.
04A judge scores the brief for grounding and coverage, then decides pass or another loop.
05Write the verified brief and its citations back to memory for the next design task.

05Tooling

Simulation, solvers & frameworks.

Get started

From engineering problems to autonomous solutions.

Talk to our engineers john.ferderigos@agentriq-sa.com