Research

We read the research, so your build starts from what works today

Anthropic and OpenAI publish what their models can do and how they build with them. We read all of it, keep up as it changes every few months, and bring what holds up into every client engagement.

Frontier models

AI now creates new knowledge

In the past months, models from OpenAI and Anthropic have settled problems that mathematicians and physicists worked on for decades. The results are published and were checked by leading researchers.

The bottleneck is no longer intelligence. It is deployment.

Disproving a famous conjecture or finishing a calculation physicists had given up on is PhD-level work, and on some problems beyond it. On exams written for PhDs, frontier models now score above the experts in the field.

So the question for a company is no longer whether the model is smart enough. It is how to bring that intelligence into everyday work: into your processes, your tools and your data, safely and measurably. Between what these models can do and what businesses use them for today lies a large overhang. Closing it is our work.

From the labs

What the engineers at Anthropic and OpenAI learned

Recent findings from the labs’ own engineering writing that matter most for a company putting AI to work, each with its figure redrawn and a link to the original.

Trusted: deterministic code
Untrusted: sandboxed
GatewayConfigures the agent, separate container
Credential brokerThe agent never sees secrets
Auto reviewA second agent approves each action
LogEvery config change and every turn
AgentIts own identity
Workspace and networkOwn directories, allowlisted hosts
configuresusesevery action

Hundreds of agents, none trusted

OpenAI has “persistent agents deployed across the entire company, including hundreds of OpenClaw agents”, and “over 500 agents that run proactively in the background doing meaningful work”. The agent and the code it writes are treated as untrusted and sandboxed; trusted, deterministic code owns networking, identity and secrets. A second agent reviews every action (“we enforce auto review on all agent actions”), and “we log everything”.

What it means for you

Agents that run on their own inside a company are safe because of the walls around them, not because the model behaves: a sandbox, a credential broker so the agent never sees secrets, its own identity, a reviewer for every action and a full log. We build those walls first.

Source: OpenAI, “Securing persistent agents like OpenClaw at OpenAI”, 12 November 2026. Our drawing of the post’s text.

Dreaming
Session
TranscriptsSearchable event log
Injected contextLoaded into every session
Memory representationsFrom transcripts and notes
Memory notesSalient observations
Next session
Promotes key notesBuilds wikisCorrects notesLoaded at start

Agents that tidy their memory offline

Anthropic splits agent memory into layers: raw transcripts, notes Claude writes during a session, derived representations such as wikis, and a small injected context loaded into every session. An offline pass called dreaming checks notes against the transcripts, corrects them and promotes what matters upward.

What it means for you

An agent that works on the same business over weeks needs memory it can trust, so a careful build keeps the raw log as ground truth and keeps the always-loaded layer small and reviewed.

Source: Anthropic, “The components of agent memory”, 6 November 2026. Redrawn from figure 1 in the post.

API cost of the cleanup

  • GPT-6 Astra Ultra142 findings$109.06
  • GPT-6.1 Sol Ultra131 findings$15.12
  • GPT-6.1 Sol Medium86 findings$5.27

A cheaper setting found 92% as much

Five OpenAI engineers ran variations of one broad cleanup prompt across repositories: “we merged more than 1,400 pull requests, deleting over 800k lines of code”, part-time over two weeks. The agent gathered its own evidence and explained each change for human review. On one task, “GPT-6.1 Sol Ultra produced 92% as many findings as GPT-6 Astra Ultra at 14% of the API cost.”

What it means for you

The biggest model at the highest effort is rarely the right default. Start small, review the agent’s work, then scale, and compare a cheaper model or effort setting on a slice of the job before paying for the expensive one everywhere.

Source: OpenAI, “Maintaining software with the GPT-6 family”, 14 October 2026. Redrawn from the post’s “Stale experiment cleanup” chart.

Pass rate

  • GPT-6 Luna, in Codex0.88
  • GPT-6 Luna, standalone loop0.74
  • GPT-5-mini, in Codex0.43
  • GPT-5-mini, standalone loop0.64

Test the tool where it will run

OpenAI and deepsense.ai evaluated the same plugin in a standalone function-calling loop and inside Codex. “Pass rates differed by 14 and 21 percentage points between the standalone loop and Codex”, in opposite directions for two models. Across three repeated runs the average barely moved, yet “23 of the 108 (21%) flipped between a pass and a fail”.

What it means for you

An agent that passes a test bench can still fail in your real setup, and a steady average can hide unstable answers. We test inside the system the agent will actually run in, repeat the runs, and have a person review a sample of answers.

Source: OpenAI and deepsense.ai, “Harness-aware evaluation of plugins”, 23 September 2026. Redrawn from table 2 in the post.

Environment: hard ceiling
External content inAudit, scan
ModelProbabilistic
External actions outAudit, scan
InputsActions

The environment sets the hard ceiling

Anthropic defends agents at three points: the content they read, the model, and the environment they run in. Model defences are strong but probabilistic, and users approved roughly 93% of permission prompts, so the environment is the hard boundary; Claude Code's OS-level sandbox cut permission prompts by 84%.

What it means for you

A careful deployment does not rely on the model or on click-through approvals alone, but limits what the agent can reach with sandboxes, file boundaries and egress controls.

Source: Anthropic, “How we contain Claude across products”, 25 May 2026. Redrawn from figure 2 in the post.

  1. Mar 4Default effort cut from high to mediumTo avoid very long latency
  2. Mar 26Cleanup of old thinking shipsA bug cleared reasoning every turn, not once
  3. Apr 7Effort default revertedUsers preferred higher intelligence by default
  4. Apr 10Thinking bug fixed (v2.1.101)Root cause took over a week to confirm
  5. Apr 16Length limits added to system promptShipped alongside Opus 4.7
  6. Apr 20Prompt reverted; all three resolvedBroader ablations showed a 3% drop (v2.1.116)
  7. Apr 23Usage limits reset for all subscribers

Three small changes, one quality drop

Reports that Claude Code had got worse traced back to three separate product changes between March 4 and April 16, each hitting a different slice of traffic on a different schedule. One bug passed multiple human and automated code reviews, unit and end-to-end tests and dogfooding, and a later ablation found a 3% drop from one system prompt line.

What it means for you

An AI product can regress without any model change, so prompt and harness changes need their own evals, soak periods and gradual rollouts.

Source: Anthropic, “An update on recent Claude Code quality reports”, 23 April 2026. Our drawing, based on the post’s text.

Tools+ resources / MCP
SessionAppend-only event log
HarnessLoop that calls Claude
SandboxRuns code, edits files
OrchestrationScheduler with retry
MCP via proxyemitEvent / getEventsexecute(name, input)wake(sessionId)

Decouple the brain from the hands

Managed Agents splits an agent into a session (an append-only event log), a harness (the loop that calls Claude) and sandboxes reached as tools through execute(name, input) → string. Any part can fail and be replaced; credentials stay in a vault behind an MCP proxy, out of reach of the sandbox.

What it means for you

Narrow, stable interfaces let a company swap models, harnesses or infrastructure later and recover from crashes without losing the agent's work.

Source: Anthropic, “Scaling Managed Agents: Decoupling the brain from the hands”, 8 April 2026. Redrawn from figure 1 in the post.

User prompt1–4 sentences
Plannerexpands into a spec
Product spec16 features, 10 sprints
Sprint contractwhat “done” means
Generatorone feature at a time
Evaluatorskeptical QA agent
Running appReact, Vite, FastAPI
writesreadsproposesreviewsfeedbackbuildsclicks through

Separate the builder from the judge

Agents asked to grade their own work tend to praise it, so Anthropic split the work across a planner, a generator and a separate evaluator that clicks through the running app. Before each sprint the generator and evaluator agree a contract on what "done" means; Sprint 3 alone had 27 criteria.

What it means for you

Agreeing testable acceptance criteria up front and having an independent check against them is the same discipline as code review and QA, applied to agents.

Source: Anthropic, “Harness design for long-running application development”, 24 March 2026. Our drawing, based on the post’s text.

Patterns we build with

How modern AI systems are actually built

This field changes every few months. These are the patterns, protocols and mental models we bring into every client engagement, drawn from what the frontier labs publish and from what holds up in production.

Legend

  • Model or agent
  • Tool or system
  • Data
  • Check
  • Person

Diagrams are our own simplified drawings of publicly documented patterns. Figures are quoted from the named source.

  • Agent patterns

    The agent loop

    An agent decides its own next step. It gathers context, acts, checks the real result and repeats, with stopping conditions and human checkpoints to keep control.

    1. Goalfrom a person
    2. Gather context
    3. Take actioncall a tool
    4. Check resultground truth
    5. Done

    Repeat until the task is done, a limit is reached or a person is needed

    Source: Anthropic, “Building effective agents”

  • Agent patterns

    Workflow or agent?

    The most useful distinction in the field. Start with the simplest thing that works and add autonomy only where it pays for its cost.

    Workflow
    • Steps follow a path written in code
    • Predictable and easy to audit
    • Best for well-defined, repeating tasks
    • Lower cost and latency
    Agent
    • The model chooses its own steps and tools
    • Flexible when steps cannot be predicted
    • Best for open-ended problems
    • Needs guardrails, testing and limits

    Source: Anthropic, “Building effective agents”

  • Retrieval and memory

    A typical RAG system

    Retrieval-augmented generation in two lanes. Documents are prepared once. Each question then pulls the most relevant passages and the model answers from them, with citations.

    IndexingDocumentsSplit into chunksEmbedSearch index
    AnsweringQuestionRetrieve passagesRerankAnswer with citations

    Source: Anthropic, “Introducing Contextual Retrieval”

  • Retrieval and memory

    Context engineering

    Everything the model sees at once competes for a limited attention budget, and recall gets worse as the window fills. The craft is choosing the smallest set of high-value tokens.

    1. System instructionsrole, rules, tone
    2. Tool definitionswhat the agent can do
    3. Retrieved dataloaded when needed
    4. Message historycompacted over time
    5. Notes and memorykept outside, read back in

    Source: Anthropic, “Effective context engineering for AI agents”

  • Tools and protocols

    Model Context Protocol (MCP)

    An open standard for connecting AI applications to other systems. The host opens one client per server, and each server offers tools, resources and prompts.

    MCP hostthe AI application
    • MCP serverfiles
    • MCP serverdatabase
    • MCP serverbusiness app
    MCP host connects to: MCP server, MCP server, MCP server.

    Source: Model Context Protocol, architecture overview

  • Safety and control

    Human in the loop

    The agent prepares, a policy decides how risky the step is, and a person approves anything that matters. Routine steps run on their own; accountability stays with people.

    AgentPrepares the actionExecutes and logs
    PolicyRates the risk
    PersonApproves or rejects
  • Quality and operations

    Evals

    You cannot improve an agent you cannot measure. Each task is run several times and scored by graders: code where possible, a model where nuance is needed, people to calibrate.

    1. Tasksfrom real failures
    2. Trialsseveral runs each
    3. Graderscode, model, human
    4. Score
    5. Regression suiteruns on every change

    Source: Anthropic, “Demystifying evals for AI agents”

Ready when you are

Which workflow should we start with?

Show us where your team loses time. In 30 minutes you will know whether AI can take it over, and what it would take to get there.